WO2010115885A1 - Predictive classifier score for cancer patient outcome - Google Patents

Predictive classifier score for cancer patient outcome Download PDF

Info

Publication number
WO2010115885A1
WO2010115885A1 PCT/EP2010/054525 EP2010054525W WO2010115885A1 WO 2010115885 A1 WO2010115885 A1 WO 2010115885A1 EP 2010054525 W EP2010054525 W EP 2010054525W WO 2010115885 A1 WO2010115885 A1 WO 2010115885A1
Authority
WO
WIPO (PCT)
Prior art keywords
tumor
patient
anatomical
parameters
images
Prior art date
Application number
PCT/EP2010/054525
Other languages
French (fr)
Inventor
Atle BJØRNERUD
Kyrre Eeg Emblem
Original Assignee
Oslo Universitetssykehus Hf
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority to EP09157255.2 priority Critical
Priority to EP09157255 priority
Application filed by Oslo Universitetssykehus Hf filed Critical Oslo Universitetssykehus Hf
Publication of WO2010115885A1 publication Critical patent/WO2010115885A1/en

Links

Classifications

    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Detecting, measuring or recording for diagnostic purposes; Identification of persons
    • A61B5/40Detecting, measuring or recording for evaluating the nervous system
    • A61B5/4058Detecting, measuring or recording for evaluating the nervous system for evaluating the central nervous system
    • A61B5/4064Evaluating the brain
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Detecting, measuring or recording for diagnostic purposes; Identification of persons
    • A61B5/74Details of notification to user or communication with user or patient ; user input means
    • A61B5/742Details of notification to user or communication with user or patient ; user input means using visual displays
    • A61B5/7435Displaying user selection data, e.g. icons in a graphical user interface
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B6/00Apparatus for radiation diagnosis, e.g. combined with radiation therapy equipment
    • A61B6/46Apparatus for radiation diagnosis, e.g. combined with radiation therapy equipment with special arrangements for interfacing with the operator or the patient
    • A61B6/461Displaying means of special interest
    • A61B6/465Displaying means of special interest adapted to display user selection data, e.g. icons or menus
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F19/00Digital computing or data processing equipment or methods, specially adapted for specific applications
    • G06F19/30Medical informatics, i.e. computer-based analysis or dissemination of patient or disease data
    • G06F19/32Medical data management, e.g. systems or protocols for archival or communication of medical images, computerised patient records or computerised general medical references
    • G06F19/321Management of medical image data, e.g. communication or archiving systems such as picture archiving and communication systems [PACS] or related medical protocols such as digital imaging and communications in medicine protocol [DICOM]; Editing of medical image data, e.g. adding diagnosis information
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/0002Inspection of images, e.g. flaw detection
    • G06T7/0012Biomedical image inspection
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/40Analysis of texture
    • G06T7/41Analysis of texture based on statistical description of texture
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Detecting, measuring or recording for diagnostic purposes; Identification of persons
    • A61B5/74Details of notification to user or communication with user or patient ; user input means
    • A61B5/742Details of notification to user or communication with user or patient ; user input means using visual displays
    • A61B5/743Displaying an image simultaneously with additional graphical information, e.g. symbols, charts, function plots
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10072Tomographic images
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/30Subject of image; Context of image processing
    • G06T2207/30004Biomedical image processing
    • G06T2207/30096Tumor; Lesion

Abstract

The present invention relates to longitudinal monitoring of tumor status and function from anatomical and functional tomographic imaging such as magnetic resonance (MR) or computed tomography (CT) images. The invention provides a new methodology for determining predictive classifiers for tumor status and function based on multi-parametric statistical classification algorithms. In addition, the invention provides a computer aided diagnostic (CAD) application encompassing several data categories and a longitudinal scrolling or browsing function providing an improved data overview tool for the user.

Description

PREDICTIVE CLASSIFIER SCORE FOR CANCER PATIENT OUTCOME

FIELD OF THE INVENTION

The present invention relates to the field of medical decision making, especially the monitoring tumors from magnetic resonance imaging (MRI), and in particular tools for longitudinal tumor monitoring and tools for predicting status and function of tumors.

BACKGROUND OF THE INVENTION

Traditionally, a combination of anatomical and functional MR or CT image information is used by experienced radiologist to assess longitudinal tumor status and growth prior to, or during treatment. In for example brain tumors (gliomas), several studies have shown good correlation between histopathological tumor grade and (a) anatomical MR image metrics and/or (b) perfusion MR image metrics prior to first time surgery - typically by injection of a MR compatible contrast agent and/or (C) diffusion MR image metrics such as fractional anisotropy. Anatomical MR images convey structural information only, acquired before, during and after contrast agent administration.

For longitudinal, life-long monitoring of untreated tumor status and post-operative residual tumor status, information of tumor size, shape and contrast uptake characteristics from MR images is widely used measures for continued patient care and in the decision making of further therapy. For assessment of treatment response from radiation therapy and/or anti-angiogenetic drug treatment, there is a movement towards use of MR perfusion and MR diffusion imaging.

Perfusion imaging based on dynamic contrast enhanced (DCE) MR or CT imaging or dynamic susceptibility contrast (DSC) MR imaging are widely used techniques for assessment of tumors grade by tumor tissue permeability (v«trans') or blood volume (BV) heterogeneity. The Ktrans permeability parameter is used to evaluate tumor cell integrity, in that the MR contrast agent leak into the extra-cellular space through a disrupted blood-brain-barrier (BBB). The blood volume is used to evaluate the micro-vascular density or vascularity, in other words, the density of small blood vessels (capillaries) in a tissue region. Traditional methods for assessment of pre- or post-operative tumor grading include manual identification of tumor regions ('hot-spots) in relevant slices by an experienced neuroradiologist. However, this method is inherently biased by user- variations - thereby elucidating the need for a CAD application with standardized criteria for measuring tumor status and function, preferably from multiple MR image parameters.

Longitudinal monitoring of tumor growth from anatomical MR images is commonly assessed by visual comparison of 2D MR images acquired at different time points. Detection of small changes in tumor size is challenging using the current method based on visual inspection alone and is further complicated by non-standardized measuring criteria, inter- and intra-observer user variations and lack of image registration of images acquired at different time-points which may make direct comparison difficult.

In addition to tumor size, the combination of patient age, neurological status, tumor composition and tumor location is often important information in tumor diagnosis. With respect to tumor location, there is high correlation between reduced overall patient survival and tumors located in eloquent locations. However, standardized, user-independent methods for assessment of tumor growth as a function of tumor location and age are not used in clinical practice.

Tumor status is traditionally classified by cell type, by grade, and/or by location. Whereas cell type and location are important parameters in the estimation of the seriousness, only the tumor grade is a predictive classifier for the patient outcome.

The WHO grading system for gliomas assigns a grade from 1 to 4, with 1 being the least aggressive and 4 being the most aggressive. Various types of astrocytomas are given corresponding WHO grades. WHO grading system for gliomas:

- WHO Grade 1 — e.g., pilocytic astrocytoma

- WHO Grade 2 — e.g., diffuse astrocytoma, oligodendrogliomas or mixed oligoastrocytoma - WHO Grade 3 — e.g., anaplastic astrocytoma, anaplastic oligodendroglioma or mixed anaplastic oligoastrocytoma

- WHO Grade 4 — glioblastoma (most common glioma in adults)

The prognosis is the worst for grade 4 gliomas, with an average survival time of less than 12 months. Overall, few patients survive beyond 3 years. Similar WHO grades are available for other types of tumors.

Presently applied grading scenarios include:

- Low-grade gliomas (WHO grades I-II) : Well-differentiated (not anaplastic); typically slow-growing tumors and portend a better prognosis for the patient.

- High-grade gliomas (WHO grades III-IV) : Undifferentiated or anaplastic; these are highly malignant associated with fast tumor growth and carry a worse prognosis.

Based on medical image information and clinical patient information (age, neurological status and current treatment regimes), a radiological tumor grade is typically assessed by a radiologist, whereas a histopathological tumor grade is assessed by a pathologist. A biopsy sample, however, has inherent limitations such as small and inaccurate tissue samples, tumor inaccessibility, subjective grading criteria, diffuse classification systems and inter- and intra-observer variability. Also, repeated invasive procedures to follow tumor recurrence, tumor growth and response to therapy are not used.

In summary, an automatic CAD application related to monitoring of tumor status and function prior to, during and after treatment with minimal user-bias would be advantageous. Also, improved tools for assessing and predicting tumor status and function would be advantageous.

A number of prior art publications describe systems for displaying medical imaging data.

• US 2008/021301 describes A method includes providing an auto visualization display based on quantitative analysis of an object 's progress over time regarding therapy response parameters over time.

• US 5,987,345 relates to a method and system for displaying medical images with computer-aided diagnosis results. A temporal subtraction tool that performs subtraction between images of a patient taken at different times to illustrate the difference in the condition of a patient over time is also described.

• US 2005/285812 relates to a method of showing medical images each representing the chest of the same human body obtained at different times in a way comparable to a "stop motion movie".

• US 2006/030768 relates to a system and method for monitoring disease progression or response to therapy using multi-modal visualization.

• US 2003/016850 discloses a systems and graphical user interface for analyzing body images, such as MR images. A marker tool allows the operator can view the statistics of the marked nodule such as volume, surface area, major and minor axes, CT density or MRI signal intensity, density and signal standard deviation, signal histogram, roundness criteria. For comparison studies, the GUI allows comparing images from a baseline scan with images from a follow-up scan. This allows the operator to assess any changes (visually and/or quantitatively) in the nodules over time.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide a system, a method and a computer program for automatically estimating and providing a predictive classifier related to pre- or post-treatment tumor status and function.

It is another object of the present invention to provide a system, a method and a computer program for automatically providing a classifier related to present glioma status and function.

It is yet another object of the present invention to provide a system, a method and a computer program for generating and presenting patient profiles with anatomical and functional tomographic image data from different acquisition dates.

Typical anatomic and functional tomographic imaging acquisitions obtains very large amounts of data, and it is practically impossible for a single person to overview and take into consideration all the information that can be extracted from this data. The present invention relates to different aspects for estimating and presenting information from post-processing of anatomic and functional tomographic imaging data, in particular to ways that serves to condense the available information to make it usable for the physician or clinician.

In this context, anatomic and functional tomographic imaging refers to techniques for generating tomograms (i.e. multiple two-dimensional image slices or sections through a three-dimensional object) with pixels/voxels representing anatomical (i.e. structural) and functional (dynamical) values. Presently, these tomographic imaging techniques are magnetic resonance (MR) imaging and computed tomography (CT) imaging.

As a result, the present invention is applicable to tumor types where functional tomographic imaging metrics have shown a high clinical efficacy, comprising intra- axial brain tumors, mammary tumors, prostate tumors, head and neck tumors, and intracranial metastasis.

In the following, reference to MR or CT imaging techniques is to be understood as a general reference to tomographic imaging techniques, which also applying to the other techniques, unless explicitly specified that the context in question can only be obtained with one of the techniques. Use of MR or CT is thereby used for purposes of brevity and exemplification only and should not be used to limit the scope of the application.

Similarly, when referring to tumor is to be understood as a general reference to, the following types: intra-axial brain tumors (e.g. gliomas), mammary tumors, prostate tumors, head and neck tumors, and intracranial metastasis, unless explicitly specified that the context in question. However, in preferred embodiments, the various aspects of the inventions are limited to intra-axial brain tumors, such as intra-axial primary brain tumors, preferably gliomas.

First aspect

In a first aspect a method, a system, and a computer program for automatically estimating and presenting a predictive classifier score (PCS) related to patient outcome for a glioma-, mammary-, prostate- or intracranial metastasis patient from anatomical and functional MR data and patient specific clinical parameters are provided. The invention solves the current technical problem with lack of a fully automated tumor classification tool with minimal user-bias. The invention solves a current problem with non-standardized criteria for tumor status, encompassing multiple parameters, such as anatomical and functional tumor parameter and clinical patient information.

Instead of a single cut-off value between tumor grades, patient groups, treatment response levels or similar, the predictive classification score of the invention will be estimated through a multi-parametric statistical classification algorithm (SCA) in a predictive classification scheme related to cancer patient outcome.

Hence, in one embodiment of the first aspect, the present invention provides a system for automatically estimating and presenting a PCS from anatomical and functional tomographic image data and patient specific clinical parameters according to claim 9.

Similarly, in other embodiments of the first aspect, the invention provides a method and a computer program product, respectively, for automatically estimating and presenting a PCS from anatomical and functional tomographic image data and patient specific clinical parameters according to claims 1 and 10.

Herein, a multi-parametric statistical classification algorithm is a procedure in which individual items (here tumor patients under examination) are placed into groups (here classifier scores) based on quantitative information on multiple characteristics inherent in the items (here the clinical patient parameters and tomographic imaging parameters) and based on a training set of items (same or other patients) previously labelled with a classifier score (here e.g. histopathological tumor grade, treatment response, time of survival, etc.). This previous labelling of the training set may be carried out retrospectively, typically from the stored patient histories in the training set. Several applicable multi- parametric statistical classification algorithms exist, and more are likely to be developed, and a few will be described in more detail herein.

In this description, a characterization or correlation to patient outcome means any label that has a demonstrated correlation to either predicted transformation from a lower tumor grade to a higher tumor grade (tumor aggressiveness), estimated time of survival, or predicted response to treatment. For the purpose of illustration and explanation, some examples of pre-defined predictive classifiers are given below. A more extended, but not exhaustive, list is presented later.

Figure imgf000008_0001

Thus, the embodiments of the first aspect of the invention rely on one or more pre-defined predictive classification schemes having classes into which the patient can be classified. Each such class is referred to as a predictive classifier score (hereafter PCS) and designate a possible patient outcome within the pre-defined scheme. Hence, that the predictive classification scheme is related to cancer patient outcome, means that the measure and classes quantifies development of cancer in patients. For the purpose of explanation of this concept, a predictive classification scheme related to, say real estate value development, could be measured by expected future price per square meter, using classes being different price intervals. Another measure could be the development in the neighbourhood using classes quantifying the degree of urbanization. The predictive classification scheme can be existing or new, future schemes.

The set of tomographic imaging parameters received from the current examination and the equivalent parameters available in the training set are not a static set of parameters. The use of multi-parametric SCA allows for the invention to be carried out on parameters sets of different composition. Hence, even though one of the preferred parameters is missing from the set, the invention may still calculate an applicable PCS. Also, which parameters are believed or demonstrated to correlate with a predictive measure may change as imaging technologies and analysis techniques evolve. The gist of the invention is to apply the described methodology of multi-parametric SCA in a predictive classification scheme related to cancer patient outcome to determine a PCS. It is therefore also preferred that the sets of training data applied by the multi- parametric statistical classification algorithm are continuously updated from new data entering a database holding the training sets.

In a preferred embodiment, the applied set of parameters comprises parameters that are presently believed to correlate with a predictive measure. These may comprises one or more of the following : a) a set of apparent diffusion coefficient (ADC) maps and/or fractional anisotropy (FA) maps from estimated from MR diffusion tensor images; b) a histogram analysis of a normalized blood volume map of an assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; c) a tumor capillary permeability map from MR perfusion images in the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; d) a tumor size determined from the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; e) a tumor location distribution from the overlap between a predefined brain region atlas and the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; and f) a tumor tissue type composition indicating the distribution of the assessed tumor region on at least the following tissue types: solid tumor tissue, edema regions, cysts, necrotic regions, and contrast enhanced regions, based on information from anatomical tomographic images and a).

The tomographic imaging parameters for the current patient, as well as the equivalent tomographic imaging parameters in the training set may be available or derivable from data in a picture archiving and communication system (PACS) hosted by a hospital or a network of hospitals.

These embodiments are particularly, but not exclusively, advantageous in that the present invention thereby provides:

- A user-independent classification, preferably based on automated determination of the MR imaging parameters used. This user independency means that the classification is much more objective that classifications with underlying parameters based on evaluation of MR imaging data of trained users.

- A more precise prediction of patient outcome than the traditional single cut-off value between tumor grades provides a better basis for decision to perform surgery and for pre-surgical planning. That the implementation of such model in an applicable tool providing a precise prediction of patient outcome is at all possible, is for the first time demonstrated herein.

- Instead of a single cut-off value between tumor grades, patient groups, treatment response levels or similar, the predictive model of the invention is utilized through a SCA applying multiple parameters that can be used to create a mode! for classification of objects based on a training set of data from previous MR examinations of one patient and/or several MR examinations from other patients with known histology and/or treatment response. A SCA can create a separating hyperplane, a higher-dimensional generalization, so that it optimally discriminates between two or more classes.

- Instead of a single parameter SCA, a multi-parametric SCA provides a more precise prediction of patient outcome. In any tumor type, no single parameter (such as tumor size, neurological status or blood volume values) can provide the physician with complete information required for an optimal treatment planning. The combination of multiple tumor parameters, however, utilized through a multi-parametric SCA procedure, provide the physician with complete, objective and 'easy-to-read' information on tumor status.

- The SCA provides a system that is continuously updated with information from new patients or new tomographic exams. Thus, instead of a stationary threshold value between tumor classes defined Once-and-for-all' during method development, the SCA relying on dynamic threshold values with a continuous improvement in diagnostic efficacy with increased size of the training set.

- The multi-parametric SCA procedure is not dependent on a constant or specific type of input data. This is advantageous since the various tumor input data types have different sizes (patient age is just one value whereas a blood volume distribution signature can have over hundred values).

- The multi-parametric SCA procedure is advantageous in that new tomographic imaging parameters can be added and redundant parameters can be removed without jeopardizing the integrity of the SCA procedure. - For each PCS, the SCA procedure provides a probability value describing the degree of certainty related to the prediction. This probability value enables a weighting system in which the prediction from each tumor class is weighted according to the degree of certainty.

The automatic estimation of a predictive classifier score is a statistical exercise performed on post-processed image data. Hence, the invention involves accessing and comparing parameters from the post-processed MR image data with equivalent parameters from previous examinations of the patient or other patients to find similarities and deviations using a multi-parametric statistical classifier algorithm. The comparison results in a predictive classifier score that is presented on a user interface, e.g. a display or a similar device, the user interface can be specific according to the application. These steps are carried out automatically by a computer accessing data on a memory and presenting a score on a display, and do not involve mental or intellectual activity of a user nor phases to be performed on the human or animal body. A previous examination phase involving acquiring MR images from patients is, however, prerequisite. It is thereby to be understood that the present invention does not provide a diagnosis as part of the estimation and presentation of a predictive classifier score. Rather, the invention provides information that can assist a physician, a clinician, and/or a technician in reaching a diagnosis or in determining a treatment.

In the following, a number of preferred and/or optional features, elements, examples and implementations will be summarized. Features or elements described in relation to one embodiment or aspect may be combined with or applied to the other embodiments or aspects where applicable. For example, functional features applied in relation to the method may also be used as features in relation to the system or computer program product and vice versa. Also, explanations of underlying mechanisms of the invention as realized by the inventors are presented for explanatory purposes, and should not be used in ex post facto analysis for deducing the invention.

The invention may encompass different sets of training data for the SCA resulting in a between-subject classification based on a single examination according to claim 2, a longitudinal between-subject classification based on several examinations according to claim 3, or a longitudinal within-subject classification based on several examinations according to claim 4.

Further, the method may apply an extra, initial single- or multi parametric statistical classification algorithm to determine an intermediate score related to patient outcome from at least one individual tomographic imaging parameter from at least the current examination, the initial statistical classification algorithm applying equivalent, previously obtained parameters as a set of training data; and thereafter using the calculated intermediate score as input to the main multi- parametric statistical classification algorithm.

The various embodiments according to the first aspect may further comprise generating a tumor report comprising o PCS from current examination o change in PCS from previous examinations o change in tumor size from previous examinations o change in tumor location and distance to functional areas from previous examinations o change in the tumor composition parameter from previous examinations and storing this in a database provided by a picture archiving and communication system (PACS).

Also, the method may further comprise a preceding post-processing of the anatomical and functional tomographic image data in the following preceding steps: i) receiving image data from functional tomographic image acquisition of the patient, comprising a. a MR diffusion tensor image acquisition; b. a perfusion image acquisition; and ii) receiving image data from an anatomical tomographic acquisition of the patient; iii) performing an automated post-processing the received images, comprising o an automated coregistering tomographic images to an anatomical image series and performing an intensity normalisation; o generation of cerebral blood volume (CBV) maps from tomographic perfusion images; o an automated tumor segmentation from the CBV maps to determine an assessed tumor region; and o an optional automated vessel segmentation to identify major vessels in the imaged regions, primarily relevant for brain and neck tumors; iv) storing the post-processed image data.

In a preferred embodiment, the invention further comprises determining an accuracy of the determined PCS and presenting it with the PCS.

The following applications of the invention may be preferred o The application of the method according to claim 2 for pre-operative prediction of patient outcome. o The application of the methods according to claims 3 and/or 4 for longitudinal pre-operative prediction of patient outcome. o The application of the method according to claim 2 for post-operative prediction of patient outcome. o The application of the methods according to claims 3 and/or 4 for postoperative prediction of patient outcome.

In a preferred embodiment, the applied multi-parametric statistical classification algorithm is a Support Vector Classification (SVM) model. In this, the contributions from feature vectors not contributing to the prediction are preferably removed or reduced by means of principal component analysis.

Second aspect

The inventors of the present invention have performed a study reported at the International Society for Magnetic Resonance in Medicine (ISMRM) 2009 Scientific meeting 2009; "MR based longitudinal assessment of pituitary adenoma growth using fully automated co-registration and intensity normalization" showing that an improved tool for generating and presentation of anatomic and functional tomographic images of a tumor from different points in time has a large influence on the assessments of the radiologist, It is shown that such tool improves the accuracy, reproducibility and objectiveness in the assessment of structural changes in the tumor compared to what can be achieved by visual comparison of images.

Presently tumor growth is commonly assessed by visual comparison of 2D MR images acquired at different time points. Detection of small changes in tumor size and appearance is challenging using visual inspection alone and is further complicated by lack of image registration of images acquired at different time- points which may make direct comparison difficult. Further, detection of functional changes in the tumor is even more difficult since assessment of functional tomographic information lack standardized criteria, is currently user-dependent and prone to inter- and intra-observer variations. Thus, the radiologist must here rely on information from previous tomographic exams, typically assessed by other radiologists, each with his or her own subjective criteria.

The use of medical imaging as a tool for gathering information provides information that can assist a physician, a clinician, and/or a technician in reaching a diagnosis is a widely used process. The efficacy of a medical imaging technique as an aid in medical diagnostic and decision making for a specific pathology, however, is not obvious and should be carefully scrutinized when developing medical imaging tools for these purposes. A detailed discussion of medical imaging efficacy can be found in Fryback and Thornbury, The Efficacy of Diagnostic Imaging, Medical Decision Making, 11, 1991, 88-94.

US 2008/021301, US 5,987,345, and US 2003/016850 disclose tools for visualizing or illustrating the difference in patient condition over time. None of these, however, suggests using the same data categories used in the embodiments of the second aspect of the invention. The prior art does also not suggest the use of a combined co-registration and intensity normalization method applied to the anatomic tomographic images or apply the same distribution analysis of functional tomographic image metrics as does the present invention. Neither is the identical functionality relating to the longitudinal selector described.

In addition, none of the above references have shown that the suggested visualizations improve the efficacy of the applied imaging techniques in medical diagnostic and decision making. In contrast, the co-registration and normalization according to the present invention has been shown to improve the efficacy of detecting changes in tumor volume compared to traditional visual assessment by expert users. A MR based study of longitudinal assessment of tumor volume change using fully automated co-registration and intensity normalization showing the improved efficacy has been reported in the ISMRM 2009 abstract referred to above and a full paper with more details and results will be published at a later stage.

In a second aspect, a new tool is proposed that makes use of a combined co- registration and intensity normalization method applied to anatomic and functional tomographic images and a new longitudinal monitoring functionality for inspecting these images.

The invention according to the second aspect solves the technical problem related to extracting and presenting anatomic and functional tomographic image data sets acquired at different time points to a user in a way that allows the user to browse the data to assess changes or non-changes in the very large amount of data.

The invention according to the second aspect preferably provides a methodology where an expert user can assess the longitudinal change in tumor size, tumor shape, tumor composition and functional tumor status (e.g. change distribution of perfusion related parameters) between different examinations through a visualization of anatomical 3D images indicating an assessed tumor region from an automatic tumor segmentation (thereby removing inter- and intra-observer variations), a graphical representation of functional tomographic imaging metrics, and relevant clinical parameters. This first aspect may be implemented as a computer aided diagnostic (CAD) application in various embodiments.

Hence, in one embodiment of the first aspect, the present invention provides a medical image evaluation system for generating and presenting a patient profile for an intra-axial brain tumor-, mammary tumor-, prostate tumor-, head and neck tumor-, or intracranial metastasis patient with anatomical and functional MR and/or CT image data from different acquisition dates according to claim 11.

Similarly, in other embodiments of the first aspect, the invention provides a method and a computer program product such as a CAD application, respectively, for generating and presenting a patient profile for an intra-axial brain tumor-, mammary tumor-, prostate tumor-, head and neck tumor-, or intracranial metastasis patient with anatomical and functional MR and/or CT image data from different acquisition dates according to claims 13 and 14.

These embodiments are particularly, but not exclusively, advantageous in that the present invention thereby aids the radiologist in making an accurate and objective assessment of structural changes of a monitored tumor. It is a further advantage that these embodiments provide the radiologist with a tool for morphometric analysis of longitudinal changes in the monitored tumor.

It is to be noted that in the present context, longitudinal is used in the meaning of something running over the course of time or being based on data or input from more than one point in time.

The distribution analysis of functional MRI metrics may for example be analysis of blood volume values, Ktrans parameters, diffusion parameters or similar.

In the present context, morphometric analysis (or morphometry) refers to assessment and visualization of structural regions in the body (such as brain cortex) and assessment and visualization of structural changes in anatomy over time (such as tumor growth).

The longitudinal selector allows the user to select a point in time and indicates the selected point in time, e.g. by moving the graphical representation f the longitudinal selector upon selection of a new point in time and by indication a selected point in time as a position on a scale or an axis. The longitudinal selector may be a scroll bar with the selection of a point in time being performed by movement of the bar (e.g. using arrow keys or drag & drop with a pointing device) and the position of the bar in the scroll panel indicating the selected point in time. Numerous of different, well known ways to graphically represent such a selecting function are known.

User interface (UI) may e.g. be a computer keyboard and or a pointing device such as touch sensitive displays, a roller mouse, scroll wheel, joystick, jog-wheel, navigation buttons etc. Further, it may be preferred that the method is applied to intra-axial brain tumors, head and neck tumors, and intracranial metastasis in which case the distribution analysis of functional MRI metrics is from the assessed tumor region less vessels determined by an automated vessel segmentation from MR perfusion images.

The configuring of retrieved data to conform to present user viewing preferences serves to enable comparison of data represented from different points in time. To be able to easily compare images or distribution analysis (such as graphs) from different acquisition dates, the consecutive images should be shown in the same way. The user viewing preferences are the values for the various data presentation parameters (e.g. zoom, rotation) which may be changed by the user, including a set of default data presentation parameters. By automatically ensuring that these user viewing preferences are the same for consecutively shown data, the longitudinal scrolling appears smooth allowing the user to immediately compare the data.

Computer-assisted diagnosis (CAD) systems are typically software applications used in medical science that supports doctors' and radiologists' interpretations and findings. Imaging techniques in e.g. MRI and CTI yield a great deal of information that the radiologist has to analyze and evaluate comprehensively in a short time. CAD systems help process digital images from the various medical imaging techniques to highlight conspicuous sections (possible diseases) and draw out relevant data for analysis and presentation. Despite their name, CAD systems do not perform a diagnosis, but simply obtains and presents relevant data in an easy comprehensible form to the doctor. It is thereby to be understood that the present invention does not provide a diagnosis as part of the presented patient profile. Rather, the invention provides information that can assist a physician, a clinician, and/or a technician in reaching a diagnosis or in determining a treatment.

In the following, a number of preferred and/or optional features, elements, examples and implementations will be summarized. Features or elements described in relation to one embodiment or aspect may be combined with or applied to the other embodiments or aspects where applicable. For example, functional features applied in relation to the method may also be used as features in relation to the system or computer program product and vice versa. Also, explanations of underlying mechanisms of the invention as realized by the inventors are presented for explanatory purposes, and should not be used in ex post facto analysis for deducing the invention.

It may be preferred that the GUI cross-fades at least the indicated assessed tumor region in the anatomical image representations upon movement of longitudinal selector to simultaneously make parts of anatomical image representations from different acquisition dates visible to the user. Such cross- fading may be implemented by morphing, transparency fading, a sliding overlapping function etc.

Similarly, it may be preferred that the GUI further comprises a cross-fading slider selector selectively cross fading the anatomical image representations by which the user simultaneously can make parts of the corresponding anatomical image representation from the temporally closest previous or succeeding acquisition dates with the present user viewing preferences visible to the user.

As described above, the co-registration, normalization of images and the configuring to conform with present user viewing preferences provides the advantage of allowing users to compare historic data at a glance. By further adding the cross-fading and/or cross-fading slider function, the invention provides the additional advantage of enabling a seamless transition between consecutively presented historic data by movement of the longitudinal selector or transparency slider. Is mentioned elsewhere, it has been shown by the inventors that improvements of this data browsing functionalities has a measurable diagnostic thinking efficacy, i.e. is able to change clinicians subjectively estimated diagnosis probabilities.

The ability to save user viewing preferences for image data for one or all acquisition dates for a patient in a picture archiving and communication systems (PACS) is a preferred further element.

Upon initial selection of a patient, retrieving all image data and clinical parameters related to the patient from the database and storing it in a random access memory (RAM). Clinical parameters are preferably available from patient data in electronic medical records (EMR) in a clinical information system (CIS).

The inventors of the present invention have performed a further study (presented at the International Society for Magnetic Resonance in Medicine, Hawaii, USA on April 21th, 2009, described in detail later herein) showing that a fully automated, user-independent predictive model for tumor status and outcome in glioma patients based on DSC MR imaging provides diagnostic accuracy values similar to those of histopathology. This is the first demonstration of an end-to-end automatic procedure for predictive characterization of gliomas, i.e. related to patient outcome.

Further aspects

In further aspects of the invention, the embodiments of the first aspect are combined with the embodiments of the second aspect to provide a system, a method or a computer program product for generating and presenting a patient profile for a glioma-, mammary-, prostate- or intracranial metastasis patient with anatomical and functional MR image data from different acquisition dates, the patient profile comprising a predictive classifier related to patient outcome automatically generated from the image data and patient specific clinical parameters.

Further embodiments of the previous aspects relates to a computer program product. Such computer program product is adapted to enable a computer system comprising at least one computer having data storage means associated therewith to control a medical image evaluation system or a unit of such to carry out the invention. Thus, it is contemplated that some known medical image evaluation systems, or units of such, may be changed to operate according to the present invention by installing a computer program product according to these aspects on a computer system controlling the medical image evaluation workstation. Such a computer program product may be provided on any kind of computer readable medium, e.g. magnetically or optically based medium, or through a computer based network, e.g. the Internet (or PACS). BRIEF DESCRIPTION OF THE FIGURES

The various aspects and embodiments of the invention will now be described in more detail with regard to the accompanying figures. The figures show one way of implementing the present invention and is not to be construed as being limiting to other possible embodiments falling within the scope of the attached claim set.

Figure 1 is an illustration of a layout of a medical image evaluation system, or of a system for estimating and presenting a predictive classifier score in accordance with various embodiments of the invention.

Figures 2-4 illustrate the methodologies behind the estimation of predictive classifier scores using a multi-parametric SCA according to embodiments of the invention.

Figures 5 and 6 illustrate presumed eloquent areas in the brain in different ways.

Figure 7 illustrates the methodology behind automated tumor segmentation as applied in embodiments of the present invention.

Figures 8 and 9 show exemplary distributions used in an example of a support vector classification model as SCA.

Figure 10 illustrates the methodologies behind the estimation of predictive classifier scores using an initial SCA for individual parameters followed by a multi- parametric SCA according to an embodiment of the invention.

Figure 11 shows Kaplan-Meier survival curves of a low-risk and high-risk group derived from actual survival data (lines 120), histopathology (lines 121) and predictive model based on DSC imaging (lines 122) over a 3 year span.

Figure 12 shows a screen shot from a CAD GUI according to an embodiment of the invention.

Figures 13A-0 and 14A-F show examples of different possible data categories and functionalities of the GUI of Figure 12 according to embodiments of the invention. DETAILED DESCRIPTION OF AN EMBODIMENT

the various embodiments, implementations, features and elements of the invention will now be described in further detail.

Figure 1 illustrates a medical image evaluation system 20 for implementing different embodiments for the various aspects of the invention. Hence, the system 20 can be a system for automatically estimating and presenting a predictive classifier score (PCS) related to patient outcome or a system for generating and presenting a patient profile with anatomical and functional tomographic image data from different acquisition dates in accordance with embodiments of the invention.

The system 20 has means 21 for receiving or accessing image data to be processed or already processed (i.e. post-processed) image data from an image recording apparatus such as a CT or MR scanner and/or internal or external storage 24 holding images recorded by such apparatus such as a PACS. The means 21 may e.g. be a data bus allowing access to a memory, an internet connection, or a cable or wireless connection. The system comprises a computer 25 or a similar processing apparatus holding an electronic processor 26 and memory 27 for holding and executing computer programs relating to the image evaluation using the received image data. The system further comprises user interface 29 (e.g. keyboard, mouse, touch screen etc.) for receiving user input related to the user viewing preferences of data presentation parameters and the longitudinal selector of the GUI. The processed image data are presented to the user via the GUI on a display 28.

Classifier Score

Figure 2 illustrates an examination course of patient A, with each dot representing Medical image acquisition and/or assessment of clinical parameters.

First, a generalized recipe for the estimation of classifiers according to the invention will be described. Thereafter, a specific example of application of a multi-parametrical SCA for derivation of a predictive classifier score will be given.

The predictive classification can be implemented in one of three forms: Method I: To give a PCS (PCSl in Figure 2) from a single examination of a patient using between-subjects comparison. Method II: To give a longitudinal PCS (PCS2 in Figure 2) for a patient using between-subjects comparison. Method II': To give a longitudinal PCS (PCS2' in Figure 2) for a patient using within-subject comparison.

Application :

For between-subjects comparisons, to compare patient information and MR metrics of a new patient with corresponding information from previous patients with known histopathological tumor grade, survival status or response to radiation-/ drug-treatment.

For within-subject comparisons, to compare patient information and MR metrics of a single patient from multiple MR examinations for longitudinal monitoring of tumor progression

Output/ result: For a new, untreated patient, it is important to assess the degree of tumor malignancy. Although the patient may undergo surgery, a biopsy sample has inherent limitation (as described elsewhere in this patent description). Thus, it is in the interest of the patient and the physician to obtain non-invasive, user- independent measures of tumor status and potential patient outcome in the form of a predictive classifier score. Other outputs can be an accuracy of the various classifiers (for instance; accuracy of survival estimate).

When : MR examinations prior to first-time surgery as well as follow-up MR examinations for residual tumor monitoring.

Goal :

At MR examinations prior to first-time surgery, the goal is to estimate the status and function of the tumor in the form of a predictive classifier such as estimated glioma grade, estimated survival time, estimated response to radiation therapy, estimated response to drug treatment. At follow-up MR examinations for residual tumor monitoring, the goal is to estimate the survival time in the form of a predictive classifier. In this, it is preferable to correct for type of surgery (biopsy/radical), type of radiation therapy (dosage) and type of drug treatment (type, dosage). 'Correct for' meaning that the multi-parametric statistical classification algorithm searches for patients in the reference database with similar characteristics (and longitudinal treatment regimes) and that the prediction is weighted towards these patients compared to other patient types.

Figure 3 illustrates the process of a 'Within-subject CAD application" of Method II' in which multiple channels of information is combined in a generalized linear classification algorithm in order to derive a patient specific report in PACS for longitudinal monitoring of tumor status and function.

Figure 4 illustrates the process of a 'Between-subject CAD application" of Method I or II in which multiple channels of information is combined in a generalized linear classification algorithm in order to derive a patient specific report in PACS prior to first time surgery.

Figures 3 and 4 serve to embody both the method according to an embodiment of the invention as well as a software architecture for the computer program product according to another embodiment of the invention. The system according to yet another embodiment can be implemented using hardware similar to that described in relation to Figure 1 previously.

Step-by-step procedure common for methods I, II, and II':

The following outlines an embodiment of the method for determining a PCS for a glioma patient. This description serves to exemplify the method and contains steps, elements and features that are not requirements for the method and also refers to MRI as an exemplary embodiment. This should not be construed as limiting the scope of the claims.

1. Image acquisition and assessment of clinical parameters The patient is at some point in time examined to obtain the tomographic image data and clinical parameters used in the multi-parametric SCA. Typical image sequences and types of clinical data applicable for this purpose will be described later in relation to the longitudinal monitoring CAD application. 2. MR image data transfer

From the MR machine, all MR image types are transferred to an online server database integrated in the hospital-wide picture and archive system (PACS). Here, it will be advantageous to obtain information from PACS on who's doing the analysis (i.e. the neuroradiologist). If so, an applied CAD application can fit the user-interface in concordance with the desired layout specific for each user (as individual default settings). For instance, if a user prefers to visualize the FLAIR images instead of the T2-w images, and prefers a coronal projection instead of axial.

3. Image post-processing

The following post-processing of acquired images is preferably performed on the acquired image data:

- From MR perfusion images: Automatic estimation of vessel segmented CBV maps and ktrans maps

- From MR diffusion tensor images: Automatic estimation of Apparent Diffusion Coefficient (ADC) maps and fractional anisotropy (FA) maps

- From BOLD-fMRI: Automatic estimation of functionally active brain areas Automatic co-registration of all MR image types to T2-weighted MR images by use of normalized mutual information or similar

Others may be used, and new may be developed as imaging technology and data analysis methods evolve.

In one embodiment, the post-processing of acquired image is carried out only as while the data is stored in the picture archiving and communication system (PACS) hosted by a hospital or a network of hospitals. Thereby, the parameters used in the present invention are available upon request.

When considering the tumor's distances to active brain areas as well as its position in relation to these, it is preferable to have a map of the relevant presumed eloquent areas in the brain to be used as a reference.

Figure 5 illustrates presumed eloquent areas in the brain shown in a sagittal projection. A tumor located within these areas is associated with lower overall survival. Al = Broca's area, A2 = precentral (motor area), A3 = postcentral gyri (sensory area), A4 = Wernicke's area, A5 = Visual cortex. The presumed location of areas A1-A5 can be estimated from BOLD-fMRI, and the analysis of tumor location should compromise information from both the brain Atlas as shown later in Figures 13J, 13K and 13L.

Figure 6 illustrates presumed eloquent areas in the brain shown in an axial MR slice through the brain (6A) and corresponding brain atlas (6B). A tumor located within these areas is associated with lower overall survival. Bl = Visual cortex, B2 = Optic radiation, B3 = Basal ganglia and internal capsule. The presumed location of area Bl can be estimated from BOLD-fMRI, area B2 from diffusion tensor imaging and B3 from the brain Atlas.

4. Automatic tumor segmentation:

From multi-spectral image analysis. A preferred automated tumor segmentation procedure is described in Emblem et al J Magn Reson Imaging. 2009 Jul;30(l) : l- 10). A shortened description is provided in the following.

A fuzzy clustering approach differ from a k-means clustering approach in that each data element (single pixel) can belong to more than one cluster thereby providing a more flexible approach than the k-mean clustering. The strength of the association between a data element and a class is indicated using a value between zero and one, in which a value closer to 1 indicate a sharper partitioning. One common fuzzy clustering approach is the FCM algorithm. This approach is based on the selection of an initial guess for the n cluster centroids, representing the mean location for each cluster. From this starting point, the cluster centroids are iteratively updated to its optimal location by minimizing an object function :

^ = ΣΣ< H ^ -CJ|2' l ≤ W ≤ ∞ [1] ι=l j=\ where u,j is the strength of the association between a given data point x, and a cluster class j, and c, is the cluster centroid. The object function in equation [1] represents the distance from x, to C1 weighted by u,j. The complementary knowledge-based operation consists of a linear sequence of low-level image processing operations based on known MR image properties secondary to brain structures or pathology. Here, for all anatomical MR image types, the glioma class was identified as the FCM cluster class with the highest image intensity. Finally, a set of binary morphological operations was performed to clean the cluster image and a 3D seed growing algorithm was applied on the complete image stack to identify and connect tumor regions in neighboring image slices. A schematic flow- diagram of the entire automatic segmentation procedure is shown in Figure 7.

In Figure 7 Schematic flow-diagram of the entire automatic segmentation procedure performed in our study. Prior to glioma segmentation, the images are standardized using adaptive histogram equalization and brain tissue pixels are identified using an intra-cranial brain mask procedure in Statistical Parametric Mapping (SPM5). The FCM cluster algorithm utilizes anatomical images (in brackets) part of a standard brain tumor MR imaging protocol. After FCM cluster analysis, a set of morphological image operations is performed on the segmented glioma images to remove non-tumor pixel areas mimicking tumor tissue.

As this segmentation procedure is somewhat complex, a potentially simpler seed- growing algorithm based on intensity values as seen on FLAIR images or T2-w images (in combination of separate) can also be used, and is considered within the scope of the invention. Here, a seed growing algorithm can be used to manually plant a seed and grow a 3D region equal to the hyper-intense T2-w area. This approach may result in a sub-optimal tumor area as a correct area should be derived from multiple 3D MR image classes (T2-w, Tl-w, FLAIR).

In a preferred implementation, steps 1 through 4 is not part of the method, but are performed previously so that the post-processed image data can be received or accessed.

5. Automatic Tumor analysis based on segmented tumor region: The automated tumor analysis serves to provide the parameters derived from functional and anatomical tomographic images to be used in the multi-parametric SCA. As described previously, these parameters accessible and actually used in the SCA may vary depending on preferences, availability and which are currently considered predictive.

In a preferred embodiment, the analysis can comprise determining one or more of: a) a set of apparent diffusion coefficient (ADC) maps and/or fractional anisotropy (FA) maps from estimated from MR diffusion tensor images; b) a histogram analysis of a normalized blood volume map of an assessed tumor region resulting from an automated tumor segmentation, optionally in combination with ktrans (i.e. CBV/ktrans matrix); c) a tumor capillary permeability map from MR perfusion images in the assessed tumor region resulting from an automated tumor segmentation; d) a tumor size determined from the assessed tumor region resulting from an automated tumor segmentation, e.g. based on real image voxel size; e) a tumor location distribution from the overlap between a predefined brain region atlas and the assessed tumor region resulting from an automated tumor segmentation, e.g. a tumor location based on normalized tumor atlas in Talairach, where the segmented tumor area is normalized to Talairach space using the T2-weighted images as reference. The resulting tumor location can then be defined in percentage: (Area 1 : 2%, Area 2: 15%, Area 3: 0 %,...,); f) a tumor tissue type composition indicating the distribution of the assessed tumor region on at least the following tissue types: solid tumor tissue, edema regions, cysts, necrotic regions, and contrast enhanced regions, based on information from anatomical tomographic images and a); g) a closest distance from tumor area to functional brain areas.

The analysis leading to each of these parameters are either well known by the person skilled in the art or described in greater detail elsewhere in this description.

In an embodiment relating to glioma and intracranial metastasis patients only, tomographic imaging parameters used in the multi-parametric SCA is derived from cerebral functional and anatomical tomographic images, and the assessed tumor region used in the histogram analysis and the tumor capillary permeability map results from an automated tumor segmentation from anatomical tomographic images less vessels determined by an automated vessel segmentation from tomographic perfusion images

In addition, further parameters can be relevant in this relation, such as a shortest distance(s) from the assessed tumor region resulting from an automated tumor segmentation from anatomical MR images to functionally active brain regions in eloquent areas (such as motor cortex, brocas and wernickes regions) determined by from BOLD-fMR images is received and included in the multi-parametric statistical classification algorithm.

6. Predictive classification using multi-parametric statistical classification algorithm

The multi-parametric statistical classification algorithm presumes a predictive classification scheme with a set of predictive classifier scores or classes into which the patient can be classified based on the MR imaging parameters and clinical patient parameters.

Several predictive classification schemes related to cancer patient outcome exist today, such as WHO glioma grade which is generally correlated to an expected survival. However, new predictive classification schemes may be developed and different ones may be used in different situations.

In the table below, a large number of possible predictive classification schemes are listed as examples with their corresponding PCS'es or classes. Some quantification of the predictive measure is typically desirable to give an easy understanding of the PCS, this is referred to as the patient outcome measure.

Figure imgf000028_0001
Figure imgf000029_0001

In the following, a more detailed description of a preferred multi-parametric statistical classification algorithm, here a Support Vector Classification Model (SVM), applying some of the above parameters is given. SVM is also a generalized linear algorithm according to a preferred embodiment of the invention. Other statistical classification algorithms are applicable, such as a k-means or fuzzy c- means clustering algorithm, algorithms based on neural networking, or logistic regression models. Some of these are also exemplified in the following.

SVM model

An example of an application SVM to glioma grading is described in the following. Say one has 10 histogram signatures as shown in Figure 8. Associated with each histogram signature is a binary value representative of glioma status from a histopathological tissue sample (high-grade = 2, low-grade = 1) :

Subject: Status:

1 2

2 1

3 1

4 1

5 1

6 2

7 2

8 1

9 1

10 2

(here, glioma grade can be replaced by survival estimates or any other relevant parameter)

Now, a neuroradiologist is interested in determine glioma grade (high-/low-grade) in a new patient with a new histogram shown in Figure 9.

In this patient, the glioma status is unknown. From the ten reference histograms in the reference database, a model is derived using SVM (or k-means, fuzzy c- means, neural networking, or logistic regression). In the SVM model (see e.g. Emblem et al., Magn Reson Med. 2008 Oct;60(4) :945-52) the model defines a 'hyperplane' between the histogram signatures with status 1 and 2. From this, the histogram signature of the new patient is compared to this hyperplane. If the histogram is 'below' the hyperplane, the SVM model predicts that the new patient has a 'high-grade' histogram signature and a glioma status value of 2 is returned. If the histogram is 'above' the hyperplane, the SVM model predict it is a low-grade glioma (=1).

In real life however, the result is often not as black/white as this example. The histogram signature can be both part 'above' and part 'below' the hyperplane. Here, the SVM model returns the most likely glioma status value (1 or 2) as previously described. In addition, it is important to pay close attention to the 'predictive accuracy value', describing how certain the SVM model prediction is. The strength of this accuracy value is related to the size of the reference database of histogram signatures. I.e., more subjects included in the reference database result in better accuracy of the SVM model.

General manuals or guides for application of SVM are described in e.g. (Boser BE, Guyon I. M., Vapnik V.N. A training algorithm for optimal margin classifiers. Proceedings of the Fifth Annual Workshop on Computational Learning Theory 1992;pp: 144-152); (Vapnik V.N. Universal Learning Technology: Support Vector Machines. NEC Journal of Advanced Technology 2005;2(2) : 137-144); or (Chang CC, Lin CJ. Training nu-support vector classifiers: theory and algorithms. Neural Comput. 2001 Sep; 13(9) :2119-2147).

An example of the use of SVM for WHO glioma grading based on a single parameter input is described in Emblem et al. Predictive modeling in glioma grading from MR perfusion images using support vector machines, Magn Reson Med. 2008 Oct;60(4):945-52, is hereby included by reference.

K-means clustering

Classification of data into sub-classes is desirable in medical imaging for separation of image regions or image metrics with similar properties, such as CBV values above or below a threshold value. In statistical analysis, cluster analysis is a common term for discrimination of data based on iterative algorithms that computes the optimal separation between a given set of classes. One simple, partitional (non-overlapping) cluster algorithm is the k-means algorithm. Here, n objects are divided into a user-specified number of cluster classes, k, of which the objects in a specific class share a common set of attributes. For image segmentation, the algorithm initially selects a set of random cluster centroids positions, and the objective of the clustering algorithm is then, through an iterative process, to minimize the within-class deviation from the class centroid and at the same time maximize the between-class centroid distance.

K-means clustering is commonly implemented as a two-step iterative process. First, based on the complete image or a smaller sub-sample, all points (i.e. image intensity values) are reassigned at once to their nearest cluster centroid, followed by a recalculation of the cluster centroids. This iteration results in an approximate solution, which reduce the computation time of the second step. Second, each data point is individually reassigned in order to assess whether the new distance reduces the sum of all distances. Although several methods exist, the distance d between the m-dimensional data point Xi, x2, , xm and a centroid position yi, y2, ym is typically assessed by deriving the squared Euclidean distance: m

1=1 [2]

Finally, the cluster centroids are recomputed until a global minimum is reached which is the optimal separation of the k cluster classes. Although a true global minimum can only be achieved by performing an exhaustive search over all starting points, a procedure with several varying staring points will generally converge towards a global minimum.

The advantage of using k-means clustering in medical imaging is that the algorithm is fast, even for a large datasets, and produce tight clusters equivalent to homogenous, compact brain tissue areas such as white or gray matter. Because a data point can belong to one cluster class only, a disadvantage of the k-means cluster algorithm is that it is relatively intolerant of imprecise data from heterogeneous image regions. Also, the number of cluster classes is an input parameter which must be selected carefully in order to avoid poor discrimination.

Logistic regression model

Logistic regression is useful for situations in which you want to be able to predict the presence or absence of a characteristic or outcome based on values of a set of predictor variables. It is similar to a linear regression model but is suited to models where the dependent variable is dichotomous. Logistic regression coefficients can be used to estimate odds ratios for each of the independent variables in the model. Logistic regression is applicable to a broader range of research situations than discriminant analysis. As an example; what lifestyle characteristics are risk factors for coronary heart disease (CHD)? Given a sample of patients measured on smoking status, diet, exercise, alcohol use, and CHD status, one can build a model using the four lifestyle variables to predict the presence or absence of CHD in a sample of patients. The model can then be used to derive estimates of the odds ratios for each factor to tell you, for example, how much more likely smokers are to develop CHD than nonsmokers.

7. Report sent back to PACS:

In order to improve the data training set, the procedure may include, as an optional step, the reporting of the assessed parameters from the estimation of a predictive classifier score.

From both Methods I, II and II', one or more of the following parameters can be stored in relation to the image data from the current examination : tumor location, distance to functional, CBV or CBV_vs_Ktrans heterogeneity, tumor size, response to therapy in similar patients, accuracy of classifier prediction. Similarly, from Method I also: predicted tumor status (grade, risk-class with respect to survival). And, from Method II also: Changes since last exam; size, location, distance to functional area, growth direction.

8. Manual/Automatic optimization routine:

It is to be preferred that the SCA is updated by an automatic, iterative optimization procedure after manual inclusion of new patients when the current information is available; tumor grade and genetic markers after MR exam and subsequent surgery/ histopathological diagnosis treatment strategies at current MR exam (none, drugs, radiation therapy) survival status

Multi-parametric SCA modelling

As described previously, the estimation of the PCS can apply an extra, initial single- or multi parametric statistical classification algorithm to determine an intermediate score related to patient outcome from at least one individual tomographic imaging parameter from at least the current examination, the initial statistical classification algorithm applying equivalent, previously obtained parameters as a set of training data; and thereafter using the calculated intermediate score as input to the main multi-parametric statistical classification algorithm.

Figure 10 illustrates an example of a 7-class tumor analysis showing the detailed deduction of the process in which a two-step statistical classification algorithm (SCA) procedure is used to derive a final prediction for a patient-specific glioma status.

In the initial step (A), the seven tumor classes from a given patient (P1-P7) is analyzed in a local (here single-parameter) SCA models (S1-S7) where the specific tumor class is compared to a local reference database (R1-R7) with corresponding data and predefined predictive classifier score (PDS). The outcome from each local SCM model (S1-S7) is a PCS value (PCS1-PCS7) and a probability value (Probability [PCS1-PCS7]) associated with each PCS value. The probability value defines the accuracy of the predicted PCS value.

In the second step (B), the PCS value and probability estimate from each local SCA model (S1-S7) is included in a global multi-parametric SCA model (Gl) to derive a final PCS value (PCS final) and a probability value (Probability [PCS Final]). In this global SCA model, the local PCS values and probability values are combined and compared to corresponding combinations of local PCS values and probability values from a reference database of patients with known PCS values.

Multi-parametrical SCA example

A specific example of application of a multi-parametrical SCA for derivation of a predictive classifier score will be given. This describes the study mentioned previously which is to be published at the International Society for Magnetic

Resonance in Medicine, Hawaii, USA on April 21th, 2009. In this description, reference numbers in square brackets refers to the following background scientific articles:

[1] Nilsen et al. Acta Oncologica 2008; 47: 1265-1270

[2] Jackson et al. AJNR Am J Neuroradiol 2002; 23 :7-14

[3] Boxerman et al. AJNR Am J Neuroradiol 2006; 27(4) :859-867 [4] Clark et al. IEEE Trans Med Imaging 1998; 17(2) : 187-201

[5] Emblem et al. Proc ISMRM 2008; p 630

[6] Emblem et al. Magn Reson Med 2008; 60(4) :945-952

The added value of using multi-parametric models encompassing such MR perfusion parameters as voxel-by-voxel cerebral blood volume (CBV) values and degree of contrast agent leakage into the extra cellular space (Ktrans,) have been shown in both DSC imaging of human gliomas [1] and dynamic contrast enhanced (DCE) MR imaging of animal tumor models [2]. Since the majority of patients suspected of a glioma will undergo surgery, pre-operative glioma grading from MR perfusion imaging may have limited value. However, current histopathological methods for grading gliomas using the World Health Organization (WHO) classification may be limited by sub-optimal criteria and sampling error. Thus, accurate prediction of patient outcome may be critical with respect to treatment planning. In this study, the inventors present a user-independent, predictive model based on analysis of CBV and κtrans parameters from DSC imaging and compare the results with patient outcome as suggested by histopathology.

In this current example, 56 adult patients (23 females, 33 males, aged 26-78 yrs, mean age 52 yrs) previously untreated patients were included in the study. Axial T2-w image, axial Tl-w images (pre- and post-contrast), coronal FLAIR images and axial gradient-echo echo-planar DSC images were acquired at 1.5 Tesla (Siemens, Germany) prior to surgery and subsequent histopathological diagnosis. CBV and Ktrans maps were derived using established methods [3] and Ktans were corrected for both "negative" (Tl-dominated) and "positive" (T2*-dominated) leakage effects. Multi-slice tumor regions were segmented automatically from the co-registered anatomical MR images using a knowledge-base fuzzy clustering technique [4] and macroscopic vessels were automatically removed based on clustering of multiple parameters of the DSC first-pass curve [5]. Based on the segmented tumor regions, 3D relative frequency scatter diagrams of CBV as a function of Ktrans for each tumor pixel were derived for each patient. This is shown in Figures 13A and B as presented previously, showing 3D surface plots of the scatter diagrams of a "high-risk" patient (13A) and a "low-risk" patient (13B). The scatter diagrams show the distribution of CBV values as a function of Ktrans for each pixel within the tumor ROI. The black histograms illustrate the distribution of CBV at Ktrans=0. The scatter diagrams were normalized, i.e. the area under the complete surface was equal to one.

The 3D scatter diagrams were transformed into a feature vector for each patient and a predictive model [6] based on support vector machines (SVM) were used to predict outcome in each patient based on the feature vectors and survival status of the remaining patients. Kaplan-Meier survival curves of a high-risk group and a low-risk group over a 3 year span were derived based on; (a) actual survival data (alive/deceased), (b) patient outcome as suggested by histology (high-/low- grade) and (c) by DSC imaging. Here, log-rank estimations with P-values were used to assess differences between the survival curves. Image analysis was performed using Matlab R2008a (MathWorks, Natick, US) and nordicICE (NordicImagingLab, Norway).

In the current example, 24 patients received a histopathological diagnosis of a low-grade glioma (WHO grade II). Of the remaining 32 high-grade patients, 7 were diagnosed with a grade III glioma and 25 were diagnosed with a grade IV glioma. Results of the mutli-parametric SCA analysis showed that the Kaplan- Meier survival curves from DSC imaging conveyed similar or better predictive values compared to histopathology shown in Figure 11. Figure 11 shows Kaplan- Meier survival curves of a low-risk and high-risk group derived from actual survival data (lines 120), histopathology (lines 121) and predictive model based on DSC imaging (lines 122) over a 3 year span. With reference to the actual survival data, the proposed model showed similar or better diagnostic accuracy values compared to histopathology. For the low-risk groups, the log-rank value between the survival curves derived from actual survival data (i.e. alive patients) and histopathology was 4.971 (P=.026). The corresponding log-rank value between actual survival data and DSC imaging was 5.46 (P=.019). For the high- risk groups, the log-rank value between the survival curves derived from actual survival data (i.e. deceased patients) and histopathology was 4.283 (P=.038). The corresponding log-rank value between actual survival data and DSC imaging was 2.901 (P=.089). The low-rank value between the high- and low-risk group derived from histopathology and DSC imaging was 17.153 (P< .001) and 19.981 (P< .001), respectively. In the current example, the inventors show the use of a user-independent approach to predict patient outcome in new patients based on DSC data from previous patients. The results suggest that the proposed method show similar or better diagnostic accuracy values to histopathology which in turn may aid in treatment planning and longitudinal monitoring of treatment response. The model is attractive in that it is user independent and that the predictive outcome is based on a 3D scatter diagram instead if a single cut-off value. In the future, information on radiation therapy, medication and molecular biomarkers should be included in the model.

In conclusion, the inventors have shown that a fully automated, user-independent method for predicting outcome in glioma patients provides similar diagnostic accuracy values to histopathology.

The individual elements of an embodiment of the invention may be physically, functionally and logically implemented in any suitable way such as in a single unit, in a plurality of units or as part of separate functional units. The invention may be implemented in a single unit, or be both physically and functionally distributed between different units and processors.

Longitudinal monitoring CAD

In the following, Figure 1 illustrates a medical image evaluation system 20 for generating and presenting a patient profile with anatomical and functional tomographic image data from different acquisition dates. The system combines data in a new way that has been shown by the inventors to improve the efficacy of detecting changes in tumor volume compared to traditional visual assessment by expert users.

Figure 12 illustrates a GUI of a CAD application for presenting a patient profile 1 with data in the various categories assigned to specific areas 2-4 and the functionalities related to the longitudinal selector 6 and the data presentation parameters 7 and 8. A more detailed explanation and examples of the various areas and functionalities for presenting data within specific categories are given in the below. Area 2

Area 2 presents a 2D or 3D representation of co-registered and intensity normalized anatomical images with an indication of an assessed tumor region resulting from an automated or semi-automated tumor segmentation from anatomical tomographic images;

Data in this area serves to indicate tumor size, shape and location, based on automatic previous tumor segmentation. In the embodiment illustrated in Figure 1, the area 2 shows a volumetric view used for qualitative overview of the tumor region or ROI. Image slices here can be visualized in all three orthogonal image projections (axial, coronal and sagittal) or as a compact 3D structure overlaid on anatomical MR images depending on user viewing preferences.

The objects in figure 12 is made out of two parts; a 3-plane orthogonal image viewer showing the axial MR image stack in its original projection (axial) and estimated projections (coronal and sagittal), and a patched volume based on a 2D stack of binary (0/1) images showing the segmented glioma volume only. In other words, the tumor volume shown in dark grey is a volume rendering of the segmented tumor region derived from the axial MR images. Alternatively, one could use a mesh routine (as in BrainVoyager) or the 3D rendering tool used in nordicICE®.

The patching process of the shown illustration was performed in Matlab: Patching is a low-level graphics function for creating patch graphics objects. A patch object is one or more polygons defined by the coordinates of its vertices as defined by the tumor segmentation. In the final view, the colouring and lighting of the patch can be specified. If the coordinate data does not define closed polygons, patch closes the polygons. The data can define concave or intersecting polygons. However, if the edges of an individual patch face intersect themselves, the resulting face may or may not be completely filled.

The patch (or mesh) encapsulate the assessed tumor region as a binary object (tumor areas=l, non-tumor areas=0). The automated tumor segmentation will be described elsewhere. Area 3

Area 3 presents a graphical representation of a distribution analysis of functional tomographic image metrics from functional MR image data in the assessed tumor region, such as graphic results from a histogram analysis of functional MRI data, typically from one or more of perfusion (DSC and DCE) MRI, diffusion MRI, BOLD- fMRI. The shown graphic results serve to give a fast understanding of the distribution of functional MR data metrics in the tumor, such as distribution of CBV values, ktrans values, ADC values or a combination of these.

In the embodiment illustrated in Figure 12, a CBV vs Ktrans matrix resulting from histogram analysis is shown in area 3 as a normalized surface plot showing CBV values as a function of Ktrans values. Two larger versions of such CBV vs Ktrans plots are shown in Figure 13A and B.

A CBV and Ktrans value is obtained for each tumor pixel as defined by the segmentation routine (binary glioma map + CBV map and ktrans map). The distribution of (CBV, κtrans) values is analysed in a 3D surface plot equal to Figures 13A and B. Here, the area under the surface plot is normalized, i.e. equal to 1. Especially, this is a sound approach when using the SVM model. However, when analysed manually by a neuroradiologist, a distribution similar to Figure 13B illustrates a low-grade (or 'low-risk') glioma patient with a homogenous distribution of CBV values (high histogram peak and narrow distribution) and ktrans values close to zero (indicating no extra-vascular leakage of contrast agent). A distribution similar to Figure 13A indicates a high-grade (or 'high-risk') glioma patient with a heterogeneous distribution of CBV values and ktrans values >0 indicating severe contrast leakage. These CBV vs Ktrans plots are just one embodiment of graphical representations of functional tomographic metrics, other representations are possible.

Area 4

Area 4 presents clinical parameters related to dynamic neurologic parameters.

Dynamic neurologic parameters such as one or more of neurological status, tumor area, tumor distance to BOLD-fMRI activation regions. This serves to give the radiologist some of the typical tumor parameters used for obtain a quick understanding of the tumor status and which are important background data for evaluating the data presented in areas 2 and 3. In the embodiment illustrated in Figure 12, shown in large scale in Figure 13C, the area 4 shows information on tumor size, predicted tumor class (this parameter may be replaced by the predictive classification scores described elsewhere), accuracy of prediction (in t- or P-values). Such values can be obtained from analysis performed by the medical image evaluation or from previous analysis embedded in DICOM file metadata.

Area 5

Area 5 presents clinical parameters related to demographic patient parameters.

Patient parameters relating to one or more of biochemical parameters such as treatment regimes (drug treatment with steroids, anti-angiogenetic agents or radiation therapy doses) and genetic tumor markers (lp/19q, p53, Ki67, MGMT); clinical parameters such as Neurological status, e.g. Karnofsky score or similar; and demographical parameters such as age and gender. These serves to provide quick insight into the biochemical, clinical and demographical parameters of the patient and are important background data for evaluating the data presented in areas 2, 3, and 4. In the embodiment illustrated in Figure 12, shown in large scale in Figure 13D, the area 5 shows information on patient name & number, scan date, and Karnofsky score. Such information can be obtained from PACS and patient data in electronic medical records (EMR) in a clinical information system (CIS).

The GUI also shows a longitudinal selector 6 for selecting a point in time from which data is to be viewed. The acquisition dates of the available anatomical and functional tomographic image data may be indicated on the longitudinal selector 6 as shown in Figure 12. The user can move the longitudinal slider 6 in the GUI by means of the user interface 29 as described in relation to Figure 1.

As described elsewhere, the GUI may be configured to cross-fade, such as by morphing or transparency, the anatomical image representations of area 2 and optionally also the graphical representation of area 3 upon movement of longitudinal selector. Thereby, parts of anatomical image representations from different acquisition dates are simultaneously visible to the user. Data presentation parameters 7 and 8 (detailed views are shown in Figures 13E and F) may be shown in a further specified area of the GUI, allowing for the user to set his/her viewing preferences. The parameters 7 and 8 are exemplary, numerous other parameters are possible such as, zoom, rotation, tilt, data projection planes in area 2, image enhancement such as brightness, contrast etc. Other data presentation parameters may affect data presented in other areas than area 2, e.g. various threshold values in image data processing which may affect the graphical representations in area 3. The user can select his/her preferred parameters in the GUI by means of the user interface 29 as described in relation to Figure 1.

In the following, a number of further optional functionalities or features of the GUI are described.

The GUI can also have a cross-fading slider 31 that can select images from different acquisition dates selectively for area 2. This is illustrated in Figure 13H. Here, the longitudinal selector is set to show image 32 in area 2. By movement of the cross-fade slider 31, however, the user can fade between the presently selected image 32 and a corresponding anatomical image representation 33 from the temporally closest previous or succeeding acquisition date, which already has the present user viewing preferences applied thereto. This allows the user with an intuitive and precise way to estimate tumor growth.

A measuring tool 34 for measuring distances between two points, or a specific functional metric value for a selected point, in the anatomical image of area 2 is shown in Figure 13H. By moving the end point using the user interface, the user can have a measure of the distance between end points or functional tomographic metric values for a point shown in the GUI. This info can be saved together with the chosen view.

It may be of interest to have a brain region atlas defined for the various tomographic image slides. Figure 131 illustrates different slices of a brain region atlas and corresponding Tl-weighted MR images, respectively. The brain atlas is defined by two radiologists in consensus based on the anatomical information in the Tl-weighted MR images. Each brain region is assigned a unique value and the right and left brain hemispheres are assigned different brain region values. Prior to brain atlas registration, the Tl-weighted MR images are normalized into a space defined by the ICBM, NIH P-20 project, and approximate that of the space described in a Talairach-Tournoux atlas. From this, the brain atlas can be applied to any other image that has been co-registered with these scans.

Figure 13J (left-to-right) illustrates binary masks (A) of the tumor region superimposed on the brain atlas (B) in order to derive location metrics for the tumor (C). In (C), the different colours represent the different brain atlas regions. This can be used to derive metric as well as in a longitudinal monitoring of tumor growth and location as described in relation to figure 13K and L below.

Figures 13K and L illustrates brain region atlases and Figure 13M a fractional anisotropy maps with superimposed assessed tumor regions and related parameters which may be shown in a further area of the GUI. Both the superimposed tumor regions and the parameters are dynamic and will thereby change upon movement of the longitudinal slider.

Figure 13K illustrates the brain region atlas with superimposed assessed tumor regions at different MR examinations (time points). Based on the segmented or manually defined tumor region, tumor area / volume and location with respect to the brain atlas are assessed. For longitudinal monitoring, variations in location as a function of time and tumor growth will be assessed.

Figure 13L illustrates the brain region atlas with superimposed assessed tumor region at different MR examinations and activations from BOLD-fMRI with different thresholds (degree of certainty in activation). The closest distance between the tumor volume in three dimensions and the motor cortex activation will be assessed. For longitudinal monitoring, variations in this distance as a function of time and tumor growth will be assessed.

Similarly, Figure 13M illustrates a fractional anisotropy maps with superimposed assessed tumor region at different MR examinations. From this, absolute FA index values in solid tumor and edematous regions can be assessed. Also, relative FA index values in tumor regions to similar regions in the healthy, contra-lateral hemisphere can be assessed. For longitudinal monitoring, variations in FA index values as a function of time and tumor growth will be assessed. Other functionalities that may be of interest are

- a mouse-scroll option to scroll between tomographic slices from the same acquisition in the anatomical image view in area 2.

- The ability to show co-registered anatomical FLAIR, Tl-w (pre- or post contrast) instead of T2-w, either separate or as overlay

- Possibility of having functional MR maps as overlay on the anatomical images (e.g. CBV maps, ADC maps or BlOOO maps, FA maps, MR multi-voxel spectroscopy)

The GUI can also present a data category being one or more graphs with a trend- lines showing dynamic data as a function of acquisition dates together with an indication of the present selection of a point in time by the longitudinal selector.

For longitudinal studies with two or more MR time points, the resulting CBV- histograms or CBV vs. ktrans matrixes can be visualized at the same time. This is illustrated in Figure 13N using a CBV-histogram 35.

Also, a trend plot 36 summarizing dynamic data (i.e. data changing over time) such as functional MR metrics or neurological parameters from different time point can be visualized as shown in Figure 130. In here, line 37 indicates the present selection of a point in time by the longitudinal selector. Line 37 will move in the trend plot upon movement of the longitudinal selector, whereas the plots themselves are static.

Longitudinal selection

Upon selection of a point in time by the longitudinal selector by the user, anatomical and functional tomographic image data acquired temporally closest to the selected point in time and clinical parameters assessed temporally closest to the selected point in time are retrieved. The retrieved data is then configured to conform with present user viewing preferences of the GUI and are presented in the relevant areas of the GUI. Figures 14A-F illustrates areas 2 and 3 for patient profiles in a longitudinal development for a patient. The development of the assessed tumor region and the functional tomographic metrics can be seen.

Data in PACS

The anatomical and functional tomographic image data used to generate the presentations in the GUI are stored in the database 24 described in relation to Figure 1. Such database is typically a picture archiving and communication systems (PACS) available at most hospitals.

This PACS preferably holds, for each patient, anatomical and functional tomographic image data acquired on different acquisition dates, the anatomical images having undergone a co-registration and intensity normalization so that all slices convey the same anatomical and geometrical information independent of acquisition date. This co-registration and normalization is described in greater detail below.

Also, assessed tumor regions resulting from an automated tumor segmentation in anatomical tomographic images from each acquisition date can be stored. The automated tumor segmentation is described in greater detail elsewhere.

In additions, which is relevant for brain and neck tumors such as gliomas, PACS can hold positions of vessels within at least the assessed tumor region resulting from an automated vessel segmentation from MR perfusion images from each acquisition date. For the distribution analysis of several functional tomographic metrics, it is of interest to remove such vessels from the assessed tumor region prior to analysis.

Co-registration and intensity normalization

A preferred co-registration routine is described in e.g. Holland et al. Proc ICAD 2008; p 249; or in Emblem KE, Holland D, Ringstad G, HaId JK, Dale AM, Bjornerud A. MR based longitudinal assessment of pituitary adenoma growth using fully automated co-registration and intensity normalization. Proc International Society for Magnetic Resonance in Medicine 2009, Hawaii, USA. Potentially, the co-registration can also be provided by alternative co-registration algorithms, e.g. implemented in nordicICE as exemplified below) or SPM. The applicability of this is to be tested through a longitudinal study on tumor growth in acusticus schwannomas. Here, longitudinal MR data (3D Tl-weighted images) is acquired from two or more time points in over 30 patients (reference to come).

Also, the co-registration is preferably a rigid body co-registration where images from different acquisitions are aligned e.g. using mathematical methods like Normalized mutual information. Alignment is maximized through maximal overlap (minimal mismatch) between the acquisitions by moving, turning or enlarging/diminishing one of the datasets. The term rigid body refers to the restriction that none of the datasets to be co-registered are allowed to change shape (i.e. no deformation of the underlying data)

Alternatively, mutual information or normalized mutual information based methods can be used to co-register the different time-points directly without prior intensity normalization. As an example, such is applied in nordicICE software distributed by NordicImagingLab (Norway). Here, an automatic inter-modality co- registration uses a mutual information based algorithm to search an optimal rigid transformation that aligns the two data sets. The implementation is based on an article by H. Sundar et. al. titled "Robust computation of mutual information using spatially adaptive meshes" . The main point in this article is to only consider a subset of the voxels in the computation, i.e. voxels that represent the non- homogenous areas. Both volumes are resampled onto the same uniform grid. This grid has the same orientation and spans the same area as the reference volume, unless functional to structural co-registration is applied. In that case, the roles are changed because functional data normally covers a smaller part of the brain compared to the structural data, and we would not want to include non- overlapping areas.

Co-registration of structural images between different time-points can also be performed using a variety of rigid-body based methods previously described in the literature. A good theoretical overview is given in : Rigid body registration. In R. S. J. Frackowiak, KJ. Friston, C. Frith, R. Dolan, KJ. Friston, CJ. Price, S. Zeki, J. Ashburner, and W. D. Penny, editors, Human Brain Function. Academic Press, 2nd edition, 2003. If the signal intensity in the different images is normalized so that the same brain structure has the same intensity at different time-point then co-registration can be performed using standard least squares optimization methods similar to those used to correct for patient motion in dynamic MRI. Finally, skull-based co-registration based on a small number of anatomical landmarks is a possible approach since the skull shape is constant over time in a given subject.

The intensity normalization serves to standardize images across scans, a standard adaptive histogram equalization procedure can possibly also be used instead, see Ashburner J, Friston K. Multimodal image co-registration and partitioning—a unified framework. Neuroimage, 1997, 6, 209-217. A second alternative is to standardize images across patients by scaling all images to a global mean value for each image slice separately.

Image acquisition and assessment of clinical parameters

The clinical parameters used to generate the applied are typically selected among the following, others may be applied as well :

1. Age

2. Gender,

3. Neurological status such as (but not exclusively) Ka rnofsky score, Neurological Performance Scale (NPS) or Sawaya Functional Grade (SFG). As a guideline, the NPS grading system is defined as follows;

- Grade 0: No neurological deficit

- Grade 1 : Some neurological deficit, but function adequate for useful work

- Grade 2: Neurological deficit causing moderate function impairment, e.g. ability to move limbs only with difficulty, moderate dysphasia, moderate paresis, some visual disturbance (e.g. field defect)

- Grade 3 : Neurological deficit causing major function impairment, e.g. inability to use limb/s, gross speech or visual disturbances

- Grade 4: No useful function, inability to make conscious responses

4. Treatment regimes and dosages (steroids, anti-angiogenetic agents or radiation therapy)

5. Genetic tumor markers such as (but not exclusively) loss of heterozygosity (LOH) on the short arm of chromosome 1 (Ip) and the long arm of chromosome 19 (19q), mutant p53 gene, Antigen identified by monoclonal antibody Ki-67 which is a cellular marker for tumor proliferation and state of the O-6-methylguanine-DNA methyltransferase (MGMT) gene which is a predictor for patient survival.

The anatomical and functional tomographic images used to generate the applied tomographic image data depends on the applied imaging technique. Below, we list relevant imaging sequences for MR and CT imaging, the applied data is typically selected from these, though others may be applied as well

Anatomical MR images:

- Tl-weighted; pre- and post-contrast images

- T2-weighted images

- FLAIR images

The anatomical MR images are preferably 3D with isotropic voxels.

Functional MR images:

- DSC and/or DCE MR perfusion images to derive e.g. CBV maps, Ktrans maps, vessel size imaging, Oxygen extraction fraction.

In this context, dynamic MR images are defined as a time series of anatomical MR images. Further, the functional MR images may comprise:

- Diffusion MR imaging (or diffusion tensor imaging) to derive ADC maps, FA maps, tensor rendering

- Susceptibility weighted images (SWI) for visualization of veins

- Multi-voxel MR spectroscopy imaging showing metabolic active areas (NAA, Choline, Creatine and Lactate)

- BOLD-fMRI showing brain areas activated after stimuli (motor areas, brocas, werniches)

Anatomical CT images:

- Pre- and post-contrast images

The anatomical CT images are preferably 3D with isotropic voxels.

Functional CT images:

- CT perfusion images to derive e.g. CBV maps and permeability (Ktrans) maps Although the present invention has been described in connection with the specified embodiments, it should not be construed as being in any way limited to the presented examples. The scope of the present invention is to be interpreted in the light of the accompanying claim set. In the context of the claims, the terms "comprising" or "comprises" do not exclude other possible elements or steps. Also, the mentioning of references such as "a" or "an" etc. should not be construed as excluding a plurality. The use of reference signs in the claims with respect to elements indicated in the figures shall also not be construed as limiting the scope of the invention. Furthermore, individual features mentioned in different claims, may possibly be advantageously combined, and the mentioning of these features in different claims does not exclude that a combination of features is not possible and advantageous.

References:

- Fryback and Thornbury, The Efficacy of Diagnostic Imaging, Medical Decision Making, 11, 1991, 88-94

- Emblem et al. Predictive modeling in glioma grading from MR perfusion images using support vector machines, Magn Reson Med. 2008 Oct;60(4):945-52

- Boser BE, Guyon I. M., Vapnik V.N. A training algorithm for optimal margin classifiers. Proceedings of the Fifth Annual Workshop on Computational Learning Theory 1992;pp: 144-152

- Vapnik V.N. Universal Learning Technology: Support Vector Machines. NEC Journal of Advanced Technology 2005;2(2) : 137-144)

- Chang CC, Lin CJ. Training nu-support vector classifiers: theory and algorithms. Neural Comput 2001 Sep; 13(9) :2119-2147)

- Nilsen et al. Acta Oncologica 2008; 47: 1265-1270

- Jackson et al. AJNR Am J Neuroradiol 2002; 23:7-14

- Boxerman et al. AJNR Am J Neuroradiol 2006; 27(4) :859-867

- Clark et al. IEEE Trans Med Imaging 1998; 17(2) : 187-201

- Emblem et al. Proc ISMRM 2008; p 630

- Emblem et al J Magn Reson Imaging. 2009 Jul;30(l) : l-10)

- Emblem et al. Magn Reson Med 2008; 60(4) :945-952

- Ashburner J, Friston K. Multimodal image coregistration and partitioning—a unified framework. Neuroimage, 1997, 6, 209-217

- Holland et al. Proc ICAD 2008; p 249 - Emblem et al. "MR based longitudinal assessment of pituitary adenoma growth using fully automated co-registration and intensity normalization", Proc International Society for Magnetic Resonance in Medicine 2009, Hawaii, USA

- Chapter: "Rigid body registration" in R. S. J. Frackowiak, KJ. Friston, C. Frith, R. Dolan, KJ. Friston, CJ. Price, S. Zeki, J. Ashburner, and W. D. Penny, editors, Human Brain Function. Academic Press, 2nd edition, 2003.

Claims

1. A method for automatically estimating and presenting a predictive classifier score (PCS) related to patient outcome for an intra-axial brain tumor-, mammary tumor-, prostate tumor-, head and neck tumor-, or intracranial metastasis patient from anatomical and functional tomographic image data and patient specific clinical parameters to be performed on suitable computer hardware, the method comprising :
receiving a set of tomographic imaging parameters derived from functional and anatomical tomographic images from a current examination;
receiving a set of clinical patient parameters for the patient relating to at least age and ongoing treatment strategies at the time of the first examination;
applying a multi-parametric statistical classification algorithm in a predictive classification scheme related to cancer patient outcome to estimate a predictive classifier score, the predictive classification scheme being based on the received clinical patient parameters and tomographic imaging parameters from at least the current examination, or scores derived there from, the multi-parametric statistical classification algorithm applying equivalent, previously obtained clinical patient parameters and tomographic imaging parameters as a set of training data; and
presenting the estimated predictive classifier score to a user.
2. The method according to claim 1, wherein the classification is a between- subject classification in that the set of training data applied by the multi- parametric statistical classification algorithm is equivalent parameters from other patients or a patient population with confirmed diagnosis and/or outcome.
3. The method according to claim 1, wherein the predictive classifier score is longitudinal in that the method further comprises receiving equivalent sets of MR imaging parameters and clinical patient parameters from the patient from one or more previous examinations;
wherein the predictive classifier score is determined based on the received parameters from the current examination and at least one previous examination; and
wherein the classification is a between-subject classification in that the set of training data applied by the multi-parametric statistical classification algorithm is equivalent parameters from other patients from at least two examinations of these.
4. The method according to claim 1, wherein the predictive classifier score is longitudinal in that the method further comprises receiving equivalent sets of MR imaging parameters and clinical patient parameters from the patient from one or more previous examination; and
wherein the classification is a within-subject classification in that the set of training data applied by the multi-parametric statistical classification algorithm is equivalent parameters from the same patient from the one or more previous examinations.
5. The method according to any of claims 1-4, further comprising the steps of: o applying at least one initial statistical classification algorithm in a predictive classification scheme related to cancer patient outcome to determine an intermediate score related to patient outcome from at least one individual MR imaging parameter from at least the current examination, the initial statistical classification algorithm applying equivalent, previously obtained parameters as a set of training data; and o using the calculated intermediate score as input to the multi-parametric statistical classification algorithm.
6. The method according to any of claims 1-5, wherein the received set of tomographic imaging parameters derived from functional and anatomical tomographic images from the current examination comprises one or more of the following : g) a set of apparent diffusion coefficient (ADC) maps and/or fractional anisotropy
(FA) maps from estimated from MR diffusion tensor images; h) a histogram analysis of a normalized blood volume map of an assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; i) a tumor capillary permeability map from MR perfusion images in the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; j) a tumor size determined from the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; k) a tumor location distribution from the overlap between a predefined brain region atlas and the assessed tumor region resulting from an automated tumor segmentation from anatomical tomographic images; and I) a tumor tissue type composition indicating the distribution of the assessed tumor region on at least the following tissue types: solid tumor tissue, edema regions, cysts, necrotic regions, and contrast enhanced regions, based on information from anatomical tomographic images and a).
7. The method according to any of claims 1-6, wherein the method relates to glioma and intracranial metastasis patients only.
8. The method according to claim 7, wherein : o the set of tomographic imaging parameters is derived from cerebral functional and anatomical tomographic images; o the assessed tumor region used in the histogram analysis and the tumor capillary permeability map results from an automated tumor segmentation from anatomical tomographic images less vessels determined by an automated vessel segmentation from tomographic perfusion images; and o a further parameter being a shortest distance(s) from the assessed tumor region resulting from an automated tumor segmentation from anatomical MR images to functionally active brain regions in eloquent areas determined by from BOLD-fMR images is received and included in the multi-parametric statistical classification algorithm.
9. A system for automatically estimating and presenting a predictive classifier score (PCS) related to patient outcome for an intra-axial brain tumor-, mammary tumor-, prostate tumor-, head and neck tumor-, or intracranial metastasis patient from anatomical and functional tomographic image data and patient specific clinical parameters, the system comprising :
computer hardware comprising an electronic processor for executing computer programs and a memory for holding computer programs;
a memory holding a computer program product enabling an electronic processor to carry out the method according to any of claims 1-8.
10. A computer program product enabling an electronic processor to carry out the method according to any of claims 1-8.
11. A medical image evaluation system for generating and presenting a patient profile for an intra-axial brain tumor-, mammary tumor-, prostate tumor-, head and neck tumor-, or intracranial metastasis patient with anatomical and functional tomographic image data from different acquisition dates, the system comprising :
- computer hardware comprising an electronic processor for executing computer programs and a memory for holding computer programs;
- a display for displaying a graphical user interface (GUI);
- a computer program held in the memory for, when executed by the processor, o accessing anatomical and functional tomographic image data for the patient from different acquisition dates; o performing an automated or semi-automated tumor segmentation; and o co-registering and intensity normalizing anatomical images;
- a computer program held in the memory for, when executed by the processor, displaying a GUI presenting at least the following data categories assigned to specific areas: o a 2D or 3D representation of the co-registered and intensity normalized anatomical images with an indication of the assessed tumor region; o a graphical representation of a distribution analysis of functional tomographic image metrics from functional MR image data in the assessed tumor region; o clinical parameters comprising dynamic neurologic parameters and demographic patient parameters the GUI further comprising : o data presentation parameters applied in the presentation of data corresponding to user viewing preferences within one or more categories; o a longitudinal selector for selecting a point in time from which data is to be presented;
- a user interface (UI) for receiving user input related to user viewing preferences and to movement of the longitudinal selector for selection of a point in time by a user; and
- computer programs for providing the following when executed by the processor: o upon selection of a point in time by the longitudinal selector, retrieving anatomical and functional tomographic image data acquired temporally closest to the selected point in time and clinical parameters assessed temporally closest to the selected point in time; and o configuring retrieved data to conform with present user viewing preferences of the GUI and providing the configured data to the GUI for presentation.
12. The medical image evaluation system according to claim 11, wherein the system further comprises access to a database provided by a picture archiving and communication systems (PACS) holding at least the following image data for each patient:
- anatomical and functional tomographic image data acquired on different acquisition dates, the anatomical images having undergone a co-registration and intensity normalization so that all slices convey the same anatomical and geometrical information independent of acquisition date;
- assessed tumor regions resulting from an automated tumor segmentation in anatomical tomographic images from each acquisition date.
13. A method for generating and presenting a patient profile for a tumor patient with anatomical and functional tomographic image data from different acquisition dates to be performed on suitable computer hardware, the method comprising : accessing anatomical and functional tomographic image data for the patient from different acquisition dates and
- performing an automated or semi-automated tumor segmentation; and
- co-registering and intensity normalizing anatomical images;
displaying a graphical user interface (GUI) presenting at least the following data categories assigned to specified areas on a display:
- a 2D or 3D representation of the co-registered and intensity normalized anatomical images with an indication of the assessed tumor region;
- a distribution analysis of functional tomographic image metrics from functional tomographic image data in the assessed tumor region;
- clinical parameters comprising dynamic neurologic parameters and demographic patient parameters; data within one or more of the categories being shown with data presentation parameters corresponding to selected user viewing preferences;
receiving user input related to selection of a point in time from which data is to be viewed, and retrieving from a database anatomical and functional tomographic image data acquired temporally closest to the selected point in time and clinical parameters assessed temporally closest to the selected point in time; and
configuring retrieved data to conform with present user viewing preferences of the GUI and presenting the configured data.
14. A computer program product enabling an electronic processor to carry out the method according to claim 13.
PCT/EP2010/054525 2009-04-03 2010-04-06 Predictive classifier score for cancer patient outcome WO2010115885A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP09157255.2 2009-04-03
EP09157255 2009-04-03

Publications (1)

Publication Number Publication Date
WO2010115885A1 true WO2010115885A1 (en) 2010-10-14

Family

ID=41559480

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2010/054525 WO2010115885A1 (en) 2009-04-03 2010-04-06 Predictive classifier score for cancer patient outcome

Country Status (1)

Country Link
WO (1) WO2010115885A1 (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013142706A1 (en) * 2012-03-21 2013-09-26 The Johns Hopkins University A method of analyzing multi-sequence mri data for analysing brain abnormalities in a subject
WO2015042421A1 (en) * 2013-09-20 2015-03-26 Siemens Aktiengesellschaft Biopsy-free detection and staging of cancer using a virtual staging score
WO2015121776A1 (en) * 2014-02-12 2015-08-20 Koninklijke Philips N.V. Systems for monitoring lesion size trends and methods of operation thereof
WO2015138385A1 (en) * 2014-03-10 2015-09-17 H. Lee Moffitt Cancer Center And Research Institute, Inc. Radiologically identifed tumor habitats
WO2016193847A1 (en) * 2015-06-03 2016-12-08 Koninklijke Philips N.V. Tumor grading using apparent diffusion co-efficient (adc) maps derived from magnetic resonance (mr) data
DE102016219607A1 (en) * 2016-10-10 2018-04-12 Siemens Healthcare Gmbh Generating a correlated with anatomical image data scorecard
EP3190542A4 (en) * 2014-09-05 2018-05-23 Universidad Politécnica de Valencia Method and system for generating multiparametric nosological images
EP2575067B1 (en) * 2011-10-01 2018-12-05 Brainlab AG Automatic treatment planning method using retrospective patient data
US10219767B2 (en) 2014-10-13 2019-03-05 Koninklijke Philips N.V. Classification of a health state of tissue of interest based on longitudinal features

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5987345A (en) 1996-11-29 1999-11-16 Arch Development Corporation Method and system for displaying medical images
US20030016850A1 (en) 2001-07-17 2003-01-23 Leon Kaufman Systems and graphical user interface for analyzing body images
US20040101181A1 (en) * 2002-07-12 2004-05-27 University Of Chicago Automated method and system for computerized image analysis prognosis
US20050108753A1 (en) * 2003-11-18 2005-05-19 Olivier Saidi Support vector regression for censored data
US20050285812A1 (en) 2004-06-23 2005-12-29 Fuji Photo Film Co., Ltd. Image display method, apparatus and program
US20060030768A1 (en) 2004-06-18 2006-02-09 Ramamurthy Venkat R System and method for monitoring disease progression or response to therapy using multi-modal visualization
US20060093198A1 (en) * 2004-11-04 2006-05-04 Fram Evan K Systems and methods for interleaving series of medical images
WO2007123914A2 (en) * 2006-04-17 2007-11-01 Siemens Medical Solutions Usa, Inc. Prognosis modeling from one or more sources of information
US20080021301A1 (en) 2006-06-01 2008-01-24 Marcela Alejandra Gonzalez Methods and Apparatus for Volume Computer Assisted Reading Management and Review
WO2008075362A2 (en) * 2006-12-20 2008-06-26 Spectrum Dynamics Llc A method, a system, and an apparatus for using and processing multidimensional data
WO2008081365A2 (en) * 2007-01-03 2008-07-10 Koninklijke Philips Electronics, N.V. Computer assisted therapy monitoring

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5987345A (en) 1996-11-29 1999-11-16 Arch Development Corporation Method and system for displaying medical images
US20030016850A1 (en) 2001-07-17 2003-01-23 Leon Kaufman Systems and graphical user interface for analyzing body images
US20040101181A1 (en) * 2002-07-12 2004-05-27 University Of Chicago Automated method and system for computerized image analysis prognosis
US20050108753A1 (en) * 2003-11-18 2005-05-19 Olivier Saidi Support vector regression for censored data
US20060030768A1 (en) 2004-06-18 2006-02-09 Ramamurthy Venkat R System and method for monitoring disease progression or response to therapy using multi-modal visualization
US20050285812A1 (en) 2004-06-23 2005-12-29 Fuji Photo Film Co., Ltd. Image display method, apparatus and program
US20060093198A1 (en) * 2004-11-04 2006-05-04 Fram Evan K Systems and methods for interleaving series of medical images
WO2007123914A2 (en) * 2006-04-17 2007-11-01 Siemens Medical Solutions Usa, Inc. Prognosis modeling from one or more sources of information
US20080021301A1 (en) 2006-06-01 2008-01-24 Marcela Alejandra Gonzalez Methods and Apparatus for Volume Computer Assisted Reading Management and Review
WO2008075362A2 (en) * 2006-12-20 2008-06-26 Spectrum Dynamics Llc A method, a system, and an apparatus for using and processing multidimensional data
WO2008081365A2 (en) * 2007-01-03 2008-07-10 Koninklijke Philips Electronics, N.V. Computer assisted therapy monitoring

Non-Patent Citations (22)

* Cited by examiner, † Cited by third party
Title
"Human Brain Function", 2003, ACADEMIC PRESS
"Human Brain Function", 2003, ACADEMIC PRESS, article "Rigid body registration"
ASHBURNER J; FRISTON K: "Multimodal image co-registration and partitioning--a unified framework", NEUROIMAGE, vol. 6, 1997, pages 209 - 217
ASHBURNER J; FRISTON K: "Multimodal image coregistration and partitioning--a unified framework", NEUROIMAGE, vol. 6, 1997, pages 209 - 217
BOSER BE; GUYON I.M.; VAPNIK V.N.: "A training algorithm for optimal margin classifiers", PROCEEDINGS OF THE FIFTH ANNUAL WORKSHOP ON COMPUTATIONAL LEARNING THEORY, 1992, pages 144 - 152
BOSER BE; GUYON I.M.; VAPNIK V.N: "A training algorithm for optimal margin classifiers", PROCEEDINGS OF THE FIFTH ANNUAL WORKSHOP ON COMPUTATIONAL LEARNING THEORY, 1992, pages 144 - 152
BOXERMAN ET AL., AJNR AM J NEURORADIOL, vol. 27, no. 4, 2006, pages 859 - 867
CHANG CC; LIN CJ: "Training nu-support vector classifiers: theory and algorithms", NEURAL COMPUT, vol. 13, no. 9, September 2001 (2001-09-01), pages 2119 - 2147
CHANG CC; LIN CJ: "Training nu-support vector classifiers: theory and algorithms", NEURAL COMPUT., vol. 13, no. 9, September 2001 (2001-09-01), pages 2119 - 2147
CLARK ET AL., IEEE TRANS MED IMAGING, vol. 17, no. 2, 1998, pages 187 - 201
EMBLEM ET AL., J MAGN RESON IMAGING, vol. 30, no. 1, July 2009 (2009-07-01), pages 1 - 10
EMBLEM ET AL., J MAGN RESON IMAGING., vol. 30, no. 1, July 2009 (2009-07-01), pages 1 - 10
EMBLEM ET AL., MAGN RESON MED, vol. 60, no. 4, 2008, pages 945 - 952
EMBLEM ET AL., MAGN RESON MED., vol. 60, no. 4, October 2008 (2008-10-01), pages 945 - 52
EMBLEM ET AL., PROC ISMRM, 2008, pages 630
EMBLEM ET AL.: "MR based longitudinal assessment of pituitary adenoma growth using fully automated co-registration and intensity normalization", PROC INTERNATIONAL SOCIETY FOR MAGNETIC RESONANCE IN MEDICINE, 2009
EMBLEM ET AL.: "Predictive modeling in glioma grading from MR perfusion images using support vector machines", MAGN RESON MED., vol. 60, no. 4, October 2008 (2008-10-01), pages 945 - 52
FRYBACK; THORNBURY: "The Efficacy of Diagnostic Imaging", MEDICAL DECISION MAKING, vol. 11, 1991, pages 88 - 94
HOLLAND ET AL., PROC ICAD, 2008, pages 249
JACKSON ET AL., AJNR AM J NEURORADIOL, vol. 23, 2002, pages 7 - 14
NILSEN ET AL., ACTA ONCOLOGICA, vol. 47, 2008, pages 1265 - 1270
VAPNIK V.N.: "Universal Learning Technology: Support Vector Machines", NEC JOURNAL OF ADVANCED TECHNOLOGY, vol. 2, no. 2, 2005, pages 137 - 144

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2575067B1 (en) * 2011-10-01 2018-12-05 Brainlab AG Automatic treatment planning method using retrospective patient data
WO2013142706A1 (en) * 2012-03-21 2013-09-26 The Johns Hopkins University A method of analyzing multi-sequence mri data for analysing brain abnormalities in a subject
US9888876B2 (en) 2012-03-21 2018-02-13 The Johns Hopkins University Method of analyzing multi-sequence MRI data for analysing brain abnormalities in a subject
WO2015042421A1 (en) * 2013-09-20 2015-03-26 Siemens Aktiengesellschaft Biopsy-free detection and staging of cancer using a virtual staging score
WO2015121776A1 (en) * 2014-02-12 2015-08-20 Koninklijke Philips N.V. Systems for monitoring lesion size trends and methods of operation thereof
US10140714B2 (en) 2014-02-12 2018-11-27 Koninklijke Philips N.V. Systems for monitoring lesion size trends and methods of operation thereof
CN106028946A (en) * 2014-02-12 2016-10-12 皇家飞利浦有限公司 Systems for monitoring lesion size trends and methods of operation thereof
WO2015138385A1 (en) * 2014-03-10 2015-09-17 H. Lee Moffitt Cancer Center And Research Institute, Inc. Radiologically identifed tumor habitats
EP3190542A4 (en) * 2014-09-05 2018-05-23 Universidad Politécnica de Valencia Method and system for generating multiparametric nosological images
US10219767B2 (en) 2014-10-13 2019-03-05 Koninklijke Philips N.V. Classification of a health state of tissue of interest based on longitudinal features
WO2016193847A1 (en) * 2015-06-03 2016-12-08 Koninklijke Philips N.V. Tumor grading using apparent diffusion co-efficient (adc) maps derived from magnetic resonance (mr) data
DE102016219607A1 (en) * 2016-10-10 2018-04-12 Siemens Healthcare Gmbh Generating a correlated with anatomical image data scorecard

Similar Documents

Publication Publication Date Title
García-Lorenzo et al. Review of automatic segmentation methods of multiple sclerosis white matter lesions on conventional magnetic resonance imaging
Shattuck et al. Online resource for validation of brain segmentation methods
Plant et al. Automated detection of brain atrophy patterns based on MRI for the prediction of Alzheimer's disease
Morton et al. Screening mammograms: interpretation with computer-aided detection—prospective evaluation
Bauer et al. A survey of MRI-based medical image analysis for brain tumor studies
Wibmer et al. Haralick texture analysis of prostate MRI: utility for differentiating non-cancerous prostate from prostate cancer and differentiating prostate cancers with different Gleason scores
van Ginneken et al. Computer-aided diagnosis: how to move from the laboratory to the clinic
Thompson et al. Cortical change in Alzheimer's disease detected with a disease-specific population-based brain atlas
Wee et al. Enriched white matter connectivity networks for accurate identification of MCI patients
Weiner et al. 2014 Update of the Alzheimer's Disease Neuroimaging Initiative: a review of papers published since its inception
Montgomery et al. Fully automated segmentation of oncological PET volumes using a combined multiscale and statistical model
Castellino Computer aided detection (CAD): an overview
US8588495B2 (en) Systems and methods for computer aided diagnosis and decision support in whole-body imaging
Leung et al. Brain MAPS: an automated, accurate and robust brain extraction technique using a template library
Peng et al. Quantitative analysis of multiparametric prostate MR images: differentiation between prostate cancer and normal tissue and correlation with Gleason score—a computer-aided diagnosis development study
Verma et al. Multiparametric tissue characterization of brain neoplasms and their recurrence using pattern classification of MR images
Gillies et al. Radiomics: images are more than pictures, they are data
US7747050B2 (en) System and method for linking current and previous images based on anatomy
Niaf et al. Computer-aided diagnosis of prostate cancer in the peripheral zone using multiparametric MRI
Vyborny et al. Computer-aided detection and diagnosis of breast cancer
US20080292194A1 (en) Method and System for Automatic Detection and Segmentation of Tumors and Associated Edema (Swelling) in Magnetic Resonance (Mri) Images
Rathore et al. A review on neuroimaging-based classification studies and associated feature extraction methods for Alzheimer's disease and its prodromal stages
Cardoso et al. LoAd: a locally adaptive cortical segmentation algorithm
Liu et al. Inter-modality relationship constrained multi-modality multi-task feature selection for Alzheimer's Disease and mild cognitive impairment identification
Lötjönen et al. Fast and robust extraction of hippocampus from MR images for diagnostics of Alzheimer's disease

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 10715169

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase in:

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 10715169

Country of ref document: EP

Kind code of ref document: A1