Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T12/00—Tomographic reconstruction from projections
  • G06T12/30—Image post-processing, e.g. metal artefact correction
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/04—Architecture, e.g. interconnection topology
  • G06N3/0464—Convolutional networks [CNN, ConvNet]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N3/00—Computing arrangements based on biological models
  • G06N3/02—Neural networks
  • G06N3/08—Learning methods
  • G06N3/09—Supervised learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N7/00—Computing arrangements based on specific mathematical models
  • G06N7/02—Computing arrangements based on specific mathematical models using fuzzy logic
  • G06N7/04—Physical realisation
  • G06N7/046—Implementation by means of a neural network
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T12/00—Tomographic reconstruction from projections
  • G06T12/10—Image preprocessing, e.g. calibration, positioning of sources or scatter correction
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/0002—Inspection of images, e.g. flaw detection
  • G06T7/0012—Biomedical image inspection
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/10—Image acquisition modality
  • G06T2207/10072—Tomographic images
  • G06T2207/10104—Positron emission tomography [PET]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/20—Special algorithmic details
  • G06T2207/20084—Artificial neural networks [ANN]
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/30—Subject of image; Context of image processing
  • G06T2207/30004—Biomedical image processing
  • G06T2207/30096—Tumor; Lesion
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2211/00—Image generation
  • G06T2211/40—Computed tomography

Definitions

• PET positron emission tomography
• Positron emission tomography (and more generally, nuclear imaging) is an important tool for early tumor diagnoses.
• Image reconstruction in PET is performed by converting coincident 511 keV photon pairs registered in a PET detector (e.g., ring) gantry into an estimate of the underlying radiotracer concentration located inside the patient.
• a PET detector e.g., ring
• the resulting visual two-dimensional or three-dimensional (2D or 3D) representation of the radiotracer distribution is used to visualize/analyse (in-vivo) biological processes linked to specific diseases.
• assessing deviations in the“normal” (i.e. physiological) tracer-accumulation (resulting in cold or hot spots) in the image is used to reveal e.g. a locally increased glucose uptake, as present in hyper- mitotic, cancerous lesions.
• Diagnostic PET imaging is thus useful to identify and quantify suspicious lesions down to the PET resolution limit (which depends on the detector geometry, crystal dimensions, positron range, sensitivity, etc.).
• the underlying physics in PET and a wide range of effects that impact image quality have been well understood and modeled in a diversity of reconstruction algorithms including compensation techniques for inelastic scattering, random coincidences, and even motion, which increases the chance of early tumor detection.
• PET imaging such as lesion standard uptake value (SUV) measurement
• SVS lesion standard uptake value
• lesions in a PET image often suffer quantification degradation due to partial volume effect.
• Resolution recovery methods based on system point spread function can be used in image reconstruction or image post-reconstruction processing to compensate the loss of lesion quantification due to partial volume effect.
• these resolution recovery methods often complicate the reconstruction algorithm, increase noise propagation in the image reconstruction and potentially create image overshoot artefact.
• these methods usually cannot fully recover the loss of lesion quantification accuracy and have varying performance from case-to-case and lesion-to-lesion, such that clinical user cannot confidently rely on the lesion quantitation measurements in the images even with resolution recovery enabled.
• the local lesion/background activity ratio and other impact factors e.g. noise regularization applied during/following image reconstruction
• these effects tend to augment the so-called partial volume effect caused by a limited spatial resolution of the scanner and/or spatial binning of the reconstructed image (i.e. the voxel/blob-size used). This leads to an increased variance in the outcome analysis not only between patients but also in subsequent studies for individuals (e.g. during treatment monitoring).
• Known approaches to compensate for SUV underestimation typically focus only on single influence factors (such as the tumor size), and typically use look-up tables generated from dedicated phantom experiments in order to provide individual correction factors.
• PET imaging devices are complex and highly configurable, typically providing features such as multiple image reconstruction modes using different image resolutions, a range of different noise reduction options, and so forth, which alone or in combination may affect a subsequent SUV assessment.
• the resulting (multi-dimensional) parameter space is complex and difficult to accommodate using existing approaches.
• related corrections recommended for inter- and intra-patient variance reduction in tumor staging and treatment monitoring may not be applied, or may be applied incorrectly leading to an incorrect quantitative SUV adjustment.
• a non-transitory computer-readable medium stores instructions readable and executable by at least one electronic processor to perform an imaging method.
• the method includes: reconstructing emission imaging data to generate an emission image of a lesion; converting intensity values of the emission image to at least one standardized uptake value (SUV value) for the lesion; processing input data using a regression neural network (NN) to output an SUV correction factor for the lesion, wherein the input data includes at least two of (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) a size of the lesion, and (iv) reconstruction parameters used in the reconstructing; and controlling a display device (24) to display at least one of (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.
• an imaging system includes an image acquisition device configured to obtain emission imaging data of a patient.
• At least one electronic processor is programmed to: reconstruct emission imaging data to generate an emission image of a lesion; convert intensity values of the emission image to at least one SUV value for the lesion; process input data using a regression neural network (NN) to output an SUV correction factor for the lesion, wherein the input data includes at least three of (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) a size of the lesion, and (iv) reconstruction parameters used in the reconstructing; generate a confidence value for the SUV correction factor by repeating the processing with one or more values of the input data varied to perform a sensitivity analysis of the SUV correction factor on the one or more values; and control a display device to display at least one of (I) the SUV correction factor; (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value; and (III) the generated confidence value.
• an imaging method includes: training a regression NN with training emission images of lesions labeled with ground truth SUV values; processing input data using the NN to output an SUV correction factor for the lesion, wherein the input data includes at least three of (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) a size of the lesion, and (iv) reconstruction parameters used in the reconstructing; and controlling a display device to display at least one of (I) the SUV correction factor and (II) a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value.
• One advantage resides in providing a more accurate lesion SUV.
• Another advantage resides in obtaining more accurate lesion SUVs without complicated and uncertain resolution recovery processing.
• Another advantage resides in providing a statistical confidence metric for the SUV of the lesion.
• Another advantage resides in allowing a user to set a reconstruction for an optimal visual image quality to detect small lesions while also obtaining an accurate quantification number for any particular lesion of interests in the image.
• Another advantage resides in providing improved SUV accuracy using a neural network efficiently trained with a large pool of clinical datasets with synthetically inserted lesions with known specifications and a mathematical model of the lesion.
• Another advantage resides in providing improved SUV accuracy using a neural network that makes fast, accurate and adaptive prediction about true lesion SUVs on the fly to a user with a normally reconstructed PET image and without the need to perform resolution recovery during or after reconstruction.
• Another advantage resides in deriving an estimate for an expected numerical mismatch in the measurement between obtained data and training data to directly correct for the SUV-max and SUV-mean values displayed to the clinician (which improves the quantitative SUV accuracy, especially in the case of small lesions), leading to more precise and reproducible results in PET -based tumor staging and therapy response assessment.
• Another advantage resides in providing variability estimations (i.e. statistical confidence values) for the displayed SUV-max and SUV-mean values, related e.g. to the specific lesion shape/size, tracer activity contrast and spatial image noise properties.
• a given embodiment may provide none, one, two, more, or all of the foregoing advantages, and/or may provide other advantages as will become apparent to one of ordinary skill in the art upon reading and understanding the present disclosure.
• FIGURE 1 diagrammatically shows an emission imaging system according to one aspect
• FIGURE 2 shows an exemplary flow chart depicting a method of operation of the system of FIGURE 1;
• FIGURE 3 shows an exemplary flow chart of a suitable approach for generating training data for the neural network of the system of FIGURE 1;
• FIGURE 4 shows another an exemplary flow chart of a suitable approach for generating training data for the neural network of the system of FIGURE 1;
• FIGURE 5 shows an exemplary flow chart depicting a method of application of the neural network of the system of FIGURE 1.
• a regression neural network is employed to generate an SUV correction factor for each lesion.
• Inputs to the network typically include a portion of the image containing the lesion and values of reconstruction parameters that are known to impact the SUV error.
• Another optional input is a delineated boundary of the lesion.
• Such inputs may alternatively be represented by salient features, for example the image may be represented by a suitable feature vector, while the lesion boundary (if serving as an input) may be represented by a lesion volume and optionally an aspect ratio. Inputting the actual image data and boundary increases the number of inputs and hence complexity of the neural network; inputting corresponding salient features reduces the needed neural network complexity at the cost of some possible reduced accuracy due to information loss.
• the neural network is a regression network which outputs the SUV correction factor (as opposed to a classification network which outputs a classification).
• Numerous neural network architectures may be employed, preferably having 3 or more layers.
• the regression neural network may, or may not, include feedback paths.
• Supervised training of the neural network receives, as training data, images of lesions with suitable ground truth labeling as to SUV values.
• experimental PET imaging data from actual patients is modified by adding "synthetic" lesions of known SUV distribution. This can be done using Monte Carlo simulation of 511 keV gamma ray emissions from the synthetic lesion at a desired position in the patient PET image, preferably adjusted for attenuation using the attenuation map (usually from CT) for the patient PET image.
• the neural network is trained (e.g., by backpropagation) on the reconstructed image converted to SUV units to optimize the network to output the correction factor adjusting to the known ground truth SUV of the synthetic lesions.
• PET imaging of a phantom with "lesions" of a priori known activity level may be used as the training samples having a priori known ground truth SUV values.
• the lesion boundary is delineated manually or by an automatic segmentation process or by a semi-automatic approach combining these, and the lesion boundary (or volume or other features derived from the boundary) serves as an input to the neural network that outputs the SUV correction factor.
• the lesion size is roughly approximated by fitting a 3D Gaussian probability distribution to the lesion, and this approximated lesion size serves as an input to the neural network.
• the lesion boundary or estimated size is not an input, and in this case the neural network may optionally be trained to output both the SUV correction factor and the lesion volume (and/or other lesion size information).
• Another optional output of the neural network is a confidence value characterizing the SUV correction factor.
• Two factors which impact this confidence are: (1) the confidence of the neural network itself; and (2) uncertainty in the lesion boundary or estimated size (for embodiments in which this is an input to the network).
• a confidence value may be generated by rerunning the neural network for designated percentage variations in the input (e.g. for lesion size, 1.05xlesion size, and 0.95xlesion size) and computing the confidence metric as the +1-5% variation values in SUV correction factor yielded for these two neural network runs.
• the neural network can be trained offline (e.g. at a factory or on-site during PET system installation or maintenance), and that the trained neural network then applied directly to clinical images to estimate the SUV correction factor.
• the neural network may be trained for different anatomical regions and types of lesions (e.g. lung lesions, prostate lesions), or a single neural network may be trained with the anatomy being a parameter input to the neural network. Since the point spread function is dependent upon the type of radiopharmaceutical, this could be another input to the neural network (or, alternatively, different neural networks may be trained for different radiopharmaceuticals).
• lesions e.g. lung lesions, prostate lesions
• a single neural network may be trained with the anatomy being a parameter input to the neural network. Since the point spread function is dependent upon the type of radiopharmaceutical, this could be another input to the neural network (or, alternatively, different neural networks may be trained for different radiopharmaceuticals).
• the output may suitably be the SUV correction factor output by the regression neural network, and/or the corrected SUV value computed by applying the SUV correction to the max- and/or average-SUV values.
• the corrected SUV value(s) are displayed along with the uncorrected SUV value(s) for consideration by the clinician.
• the confidence value (if computed) is also displayed.
• the disclosed training methods produce a regression neural network that takes into account the complex SUV dependencies on various (user dependent and independent) impact factors which are captured by a set of appropriate training data. Afterwards, when provided with the matching input from a new PET study being performed for clinical diagnosis or assessment of a patient, the trained regression neural network automatically derives an estimate for the expected numerical mismatch between the SUV value(s) of the lesion generated by the PET imaging study and the actual lesion activity (i.e., the“true” SUV).
• This mismatch estimate can be applied in order to (a) directly correct for the SUV-max and SUV- mean values displayed to the clinician (which improves the quantitative SUV accuracy, especially in the case of small lesions whose SUV values are most adversely affected by confounding factors such as partial volume effect, system blurriness, and noise.
• variability estimations i.e. confidence values
• a deep learning neural network can help to link between the observed lesion quantification (e.g. SUV measured and the true lesion quantification, i.e., true SUV) by learning from a large pool of training data about the partial volume effect of the imaging chain.
• the observed lesion quantification e.g. SUV measured and the true lesion quantification, i.e., true SUV
• this approach may have a feasibility issue that the mapping may not be a one- to-one in reality, when the true lesion size is unknown and especially when the lesion size is comparable to the system’s partial volume kernel size.
• Lesions with different SUV and different sizes can result in a similar observed lesion with the same partial volume effect applied (e.g., a narrow but tall lesion may end up similar to a wide and short lesion).
• the SUV can be measured accurately which makes the NN prediction less useful.
• the disclosed approaches can be disclosed in other emission imaging modalities in which a radiopharmaceutical is administered to a patient, such as single photon emission computed tomography (SPECT) imaging systems, hybrid PET/CT or SPECT/CT imaging systems, and the like.
• SPECT single photon emission computed tomography
• hybrid PET/CT or SPECT/CT imaging systems and the like.
• the system 10 includes an image acquisition device 12.
• the image acquisition device 12 comprises a PET gantry of a PET/CT imaging system that also includes a computed tomography (CT) gantry 13.
• CT computed tomography
• a patient table 14 is arranged to load a patient into an examination region 16 of the PET and/or CT gantries 12, 13.
• the PET gantry 12 includes an array of (diagrammatically indicated) radiation detectors 17.
• the system 10 also includes a computer or workstation or other electronic data processing device 18 with typical components, such as at least one electronic processor 20, at least one user input device (e.g., a mouse, a keyboard, a trackball, a dictation microphone for dictating a radiology report, and/or the like) 22, and a display device 24.
• the display device 24 can be a separate component from the computer 18.
• the at least one electronic data processing device 18 includes a first electronic data processing device 18i which serves as an imaging device controller (e.g. a PET scanner controller) and a second electronic data processing device I8 2 which serves as a radiology workstation.
• a radiology technician or other medical professional operates the PET scanner 12 using the PET controller I8 1 to acquire PET images, and to convert to SUV units (e.g. SUV images and/or average or maximum SUV values for lesions).
• SUV units e.g. SUV images and/or average or maximum SUV values for lesions.
• PACS Picture Archiving and Communication System
• the PACS may go by another nomenclature such as a Radiology Information System, RIS, or so forth.
• the image processing also employs a neural network (NN) 28 to correct the SUV values as disclosed herein.
• the NN 28 can be a regression NN that is trained to determine an SUV correction factor for a lesion in images acquired by the PET gantry 12.
• Data can be input to the NN 28, including (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) a size of the lesion, and (iv) reconstruction parameters used in the reconstructing; (v) an identification of a radiopharmaceutical used in acquiring the emission imaging data; among others.
• This input data can be processed by the NN 28 to determine the SUV correction factor.
• a radiologist can operate the radiology workstation I8 2 to perform a reading of the PET images, including retrieving (from the PACS 26) and comparing PET images from the current PET study and a previous PET study.
• the previous PET study may have been performed before commencement of chemotherapy, radiation therapy, or other oncology therapy, while the current PET study may have been performed after such therapy.
• the previous and current PET studies may have been performed at different times during the ongoing fractionated therapy.
• each of the PET controller I8 1 and the radiology workstation I8 2 include one or more display devices 24; the illustrative radiology workstation I8 2 includes an illustrative two displays 24, e.g.
• the radiologist may invoke an SUV correction estimate employing the neural network 28 to correct the SUV values, or to provide a proposed SUV correction. While in this illustrative example the neural network 28 is applied at the radiology workstation I8 2 , in other embodiments it may be applied at the PET controller I8 1 as an adjunct to the PET image reconstruction, so that the corrected SUV values or the proposed SUV correction factor is uploaded to the PACS 26 with the imaging study for consideration by the radiologist reviewing the study at the radiology workstation I8 2 .
• the at least one electronic processor 20 is operatively connected with the one or more non-transitory storage media (not shown; such as a magnetic disk, RAID, or other magnetic storage medium; a solid state drive, flash drive, electronically erasable read-only memory (EEROM) or other electronic memory; an optical disk or other optical storage; various combinations thereof; or so forth) which stores instructions which are readable and executable by the at least one electronic processor 20 to perform disclosed operations including performing an imaging method or process 100.
• various portions of the imaging method or process 100 are performed by a technician using the PET controller I8 1 and/or by a radiologist operating the radiology workstation I8 2 , and may be performed at least in part by cloud processing.
• an illustrative embodiment of the imaging method 100 is diagrammatically shown as a flowchart.
• the image acquisition device 12 e.g., the PET imaging device
• the at least one electronic processor 20 specifically the PET controller I8 1 in the illustrative example of FIGURE 1
• the PET controller I8 1 is programmed to reconstruct the PET imaging data into PET emission images of a lesion.
• the PET controller I8 1 is programmed to converting intensity values of the PET emission image to at least one SUV value for the lesion. (Alternatively, this could be done at the radiology workstation I8 2 ).
• the SUV value could be, for example, a maximum SUV value anywhere in the lesion; an average SUV value over the lesion, or some other representative SUV value that represents the activity of the lesion as a whole. This statistical value (e.g. max and/or average SUV) could then be corrected using the SUV correction factor. It is also the case that in the SUV image, each voxel (both inside and outside the lesion) can have an SUV value. In other words, the original intensity image is converted to an SUV image by converting the value of each voxel to an equivalent SUV value.
• the converting of intensity values of the emission image to at least one SUV value for the lesion includes converting an intensity value 7 to a corresponding SUV value according to Equation (1): where i is the index of a voxel of the PET image, v t is the value of the voxel i (expressed as a radiotracer activity concentration in the tissue at voxel i, e.g.
• the at least one SUV value can be determined by operations described in, for example, (1) Boellaard, Standards for PET Image Acquisition and Quantitative Data Analysis, The Journal of Nuclear Mediccine, Vol. 50, No. 5 (Suppl), May, 2009, pages 11S-20S; and (2). Kinahan, et al. PET/CT Standardizaed Uptake Values (SUVs) in Clinical Practice and Assessing Response to Therapy, Semin Ultrasound CT MR. December 2010; 31(6): 496-505.
• the PET controller 18i (or, alternatively, the radiology workstation I8 2 , or alternatively a cloud computing resource or other remote computer(s) operatively connected with one of these systems I8 1 , 18 2 ) is programmed to processing input data using the regression NN 28 to output an SUV correction factor for the lesion.
• the input data can include at least two of (i) image data comprising the emission image or a feature vector representing the emission image, (ii) the at least one SUV value, (iii) a size of the lesion, and (iv) reconstruction parameters used in the reconstructing.
• the input data can include at least three of these options.
• the input data can further include (v) an identification of a radiopharmaceutical used in acquiring the emission imaging data.
• the size of the lesion can be represented as a delineated boundary of the lesion in the emission image.
• the radiology workstation I8 2 is programmed to receive a manual delineating of the lesion boundary, e.g. using a manual feature contouring graphical user interface (GUI) of a type used to delineate targets and organs-at-risk in radiation therapy planning and like applications.
• GUI graphical user interface
• the radiology workstation I8 2 is programmed to delineate the lesion boundary by automatic segmentation of the emission image, or by fitting a Gaussian volume to the lesion, or so forth.
• the input data does not include the size of the lesion.
• the NN 28 is trained to output a boundary of the lesion and/or a volume of the lesion.
• the volume of the lesion can be determined by fitting a three-dimensional Gaussian probability distribution to the lesion in the emission imaging data.
• the PET controller 18i (or, alternatively, the radiology workstation I8 2 , or alternatively a cloud computing resource or other remote computer(s) operatively connected with one of these systems I8 1 , I8 2 ) is programmed to optionally generate a confidence value for the SUV correction factor.
• the PET controller I8 1 is programmed to apply the NN 28 to repeat processing of the input data with one or more values of the input data varied to perform a sensitivity analysis of the SUV correction factor on the one or more values.
• the size of the lesion is varied to perform the sensitivity analysis of the SUV correction factor on the size of the lesion.
• the PET controller I8 1 (or, alternatively, the radiology workstation I8 2 , or alternatively a cloud computing resource or other remote computer(s) operatively connected with one of these systems I8 1 , I8 2 ) is programmed to control the display device 24 to display the SUV correction factor.
• a corrected SUV value generated by applying the SUV correction factor to the at least one SUV value may also (or alternatively) be displayed on the display device 24.
• other parameters that are determined e.g., the boundary and volume of the lesion, the confidence value, and so forth can be displayed on the display device 24.
• an illustrative embodiment of a method 200 for generating training data for training the NN 28 is diagrammatically shown as a flowchart.
• the NN 28 is trained with training emission images of lesions labeled with ground truth SUV values.
• Monte Carlo simulations of synthetic gamma ray emissions are performed from a synthetic lesion inserted into training emission imaging data.
• a combination of the synthetic gamma ray emissions and the training emission imaging data is reconstructed to generate one of the training emission images which is of the synthetic lesion.
• the NN 28 is programmed to predict true SUV values.
• a mathematical model e.g. a 3D Gaussian model
• the NN 28 is then trained using a large number of patient data sets with different patient specs, acquisition time, duration time, statistics and so forth with simulated inserted lesions with known size and SUV.
• the output of the network is the blurring kernel at different spatial location in the image given different imaging specifications.
• the blurring kernel is predicted at the location of the lesion with the given imaging specifications.
• the 3D Gaussian model is generated based on the profile of the lesion observed in the image.
• the true SUV of the lesion is predicted using the blurring kernel and the measured lesion profile model.
• the lesion profile can be modelled based on a Gaussian and total intensity of the lesion measured on the image, and/or based on the lesion size measured on the CT image.
• the NN 28 can be trained using training data in which volumetric radiotracer distributions are reconstructed from Monte Carlo-simulated (or measured) PET scan data, considering, for a given PET scanner geometry, a variety of different lesion configurations.
• volumetric radiotracer distributions are reconstructed from Monte Carlo-simulated (or measured) PET scan data, considering, for a given PET scanner geometry, a variety of different lesion configurations.
• adjustments accessible to the clinicians changes accessible to the clinicians (changes to the reconstruction, noise-regularization or filtering parameters) are emulated.
• the regional SUV values are determined using the known lesion mask.
• Given the ground truth and image-based SUV measures, matching correction factors are determined.
• an Artificial Neural Network (ANN) algorithm is trained to learn the relationship between these provided inputs.
• ANN Artificial Neural Network
• FIGURE 4 shows an example method 300 of training the NN 28 depicted as a flow chart.
• ground truth data are generated using a particle simulation framework capable to adequately model the process from local positron decay to final coincidence photon detection, such as GEANT4/GATE.
• the geometrical model considers relevant scanner properties (e.g. detector module setup) of the targeted PET system.
• the scanner/detector model is combined with a lesion-background-model (“phantom”), considering, among other factors, different (e.g. radial) lesion positions inside the scanners Field-o f- View, different lesion extensions (i.e.
• each list-mode data set is subsequently reconstructed using the targeted product reconstruction.
• an (application typical) range of clinically available parameter adjustments are applied (all resulting in individual reconstruction results), including different spatial sampling; image resolution settings; different configurations regarding subset size and iteration number; different parameter settings related to image pre-/post-filtering; different configurations affecting (e.g. intrinsic) noise-regularization; and so forth.
• the reconstructed activity distributions are (point-wise) multiplied with the corresponding sensitivity map and weighted with the acquisition time in order to derive an estimate for the spatial distribution of detected decay events.
• the output is spatially cropped around a lesion and (if required) resampled in order to create a standardized, predefined input format for the NN 28.
• Each input data set further considered is assumed to specify only one lesion.
• the individual lesion definition mask is applied on the reconstructed image data in order to extract the accumulated, regionally recovered tracer amount.
• the image-derived lesion SUV SUVi
• SUV-calibration information extracted from an existing scanner or an average value from multiple scanner calibrations may be applied.
• a regional ground truth SUV value (SUVg) is calculated for the individual lesion from the known phantom setup.
• SUV-calibration information extracted from an existing scanner or an average value from multiple scanner calibrations may be applied.
• an SUV correction factor is calculated from 308 and 310 as SUVg/SUVi.
• the mapping between the lesion samples and the individual SUV correction factors is trained (e.g. via backpropagation ) to a multilayer-NN.
• Modem NN architectures applied in image processing such as (deep) Convolutional Neural Networks (CNNs), typically consist of a stack of individual layers, including convolution layers (convolution-based filtering/feature enhancement), max-pool-layers (dimensionality reduction), fully-connected layers (feature re-combination), etc. They also include additional layers/techniques (such as drop-out and regularization ) in order to preserve the ability of ANNs to abstract the feature space based on the seen training examples. Advanced techniques like transfer learning or fine tuning which allow to include public-available pre-trained ANN- layers into a dedicated specific NN could also be considered.
• the (iterative) NN weight adaptation can be performed on a single computer or a compute cluster, either using CPU or GPU.
• Different sources of public available NN modeling toolkits exist (TensorFlow, Torch, Caffe/Caffe2, Keras, and so forth) which allow to create various types of NN structures.
• the NN 28 can be used to output the SUV correction factor.
• the NN 28 responds with SUV correction factors based on the learned parameter space.
• the correction factor is either directly applied to compensate for the estimated SUV error, or it is displayed along with the un-corrected lesion SUV for subsequent (manual) correction by the clinician. Additional information about the local variability of the estimated SUV correction factors are determined, e.g. via small-range variations in selected parameters (such as the size and-or shape of the lesion segmentation).
• the SUV variability estimate is (graphically or numerically) displayed along with the corrected or un-corrected SUV evaluation results in order to indicate the“trust level” for the related analysis.
• FIGURE 5 shows an example method 400 of using the trained NN 28 depicted as a flow chart.
• a manual or semi-automatic lesion delineation for the given PET study is performed by the clinician.
• a weighting and lesion-centric regional cropping and resampling is applied.
• a lesion sub-volume together with the reconstruction/acquisition parameter set is fed into the NN 28, which derives an estimate for an appropriate SUV correction factor via a multi-dimensional projection of the provided input into the learned parameter space.
• minor (random) shape variations are applied on the used segmentation mask (and also other parameters). The resulting changes in the SUV correction factor output of the NN 28 allows to estimate the reliability of this parameter.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Software Systems (AREA)
• General Health & Medical Sciences (AREA)
• Health & Medical Sciences (AREA)
• Computing Systems (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• Biomedical Technology (AREA)
• Molecular Biology (AREA)
• Artificial Intelligence (AREA)
• General Engineering & Computer Science (AREA)
• Mathematical Physics (AREA)
• Life Sciences & Earth Sciences (AREA)
• Biophysics (AREA)
• Computational Linguistics (AREA)
• Automation & Control Theory (AREA)
• Fuzzy Systems (AREA)
• Algebra (AREA)
• Computational Mathematics (AREA)
• Mathematical Analysis (AREA)
• Mathematical Optimization (AREA)
• Pure & Applied Mathematics (AREA)
• Medical Informatics (AREA)
• Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
• Radiology & Medical Imaging (AREA)
• Quality & Reliability (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Nuclear Medicine (AREA)

Priority Applications (4)

Applications Claiming Priority (3)

Publications (2)

Family

ID=68501620

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (12)

Citations (5)

Family Cites Families (16)

• 2019
  • 2019-11-08 US US17/292,044 patent/US12346998B2/en active Active
  • 2019-11-08 JP JP2021523746A patent/JP7359851B2/ja active Active
  • 2019-11-08 WO PCT/EP2019/080628 patent/WO2020099250A1/en not_active Ceased
  • 2019-11-08 EP EP19801009.2A patent/EP3881289A1/en not_active Withdrawn
  • 2019-11-08 CN CN201980081881.5A patent/CN113196340B/zh active Active

Patent Citations (5)

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events