WO2025106981A1 - Systems and methods for localized brain damage prognosis - Google Patents

Abstract

A method for analyzing brain imaging data from a brain lesion patient and generating prognostic information is disclosed. The method includes obtaining imaging data of the brain of a patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest. The method further involves processing the data series of interest to generate brain masks and corresponding volumetric data. The brain masks and corresponding volumetric data are then co-registered to a common brain space, and registered volumetric data is generated. Lesion segmentation is performed to generate a lesion mask in the common brain space. The lesion mask data is applied to a trained prognostic model, which comprises a machine learning model configured to output a probability-based prognosis of the patient's brain lesion.

Description

SYSTEMS AND METHODS FOR LOCALIZED BRAIN DAMAGE PROGNOSIS

CROSS-REFERENCE TO RELATED APPLICATION(S)

[0001] This application claims the benefit under 35 U.S.C. § 119(e) to U.S. Provisional Application 63/600,158, filed November 17, 2023, and entitled “Personalized Prognosis from Brain Imaging” which is hereby incorporated herein by reference in its entirety.

GOVERNMENT SUPPORT

[0002] This invention was made with government support under NS 114405 awarded by the National Institutes of Health. The government has certain rights in the invention.

BACKGROUND

[0003] Stroke stands as a formidable challenge in healthcare, with its aftermath often leaving individuals facing a spectrum of cognitive impairments that can severely affect their quality of life, autonomy, and longevity. The unpredictable nature of post-stroke recovery trajectories accentuates the critical need for innovative prognostic tools capable of offering early and accurate predictions. Such advancements have the potential to aid in early intervention decisions, enhance the customization of rehabilitation plans, and aid healthcare providers and families in anticipating the support needed throughout the recovery process. However, available options are lacking and/or rely on time-consuming manual image processing and analysis.

[0004] Therefore, there is the need for an automated and highly accurate method for establishing a patient’s prognosis following a brain lesion.

[0005] While multiple embodiments are disclosed, still other embodiments of the disclosure will become apparent to those skilled in the art from the following detailed description, which shows and describes illustrative embodiments of the disclosed compositions, systems and methods. As will be realized, the disclosed compositions, systems and methods are capable of modifications in various obvious aspects, all without departing from the spirit and scope of the disclosure. Accordingly, the drawings and detailed description are to be regarded as illustrative in nature and not restrictive.

BRIEF SUMMARY [0006] In one general aspect, method may include obtaining imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest. Method may also include processing the data series of interest to generate brain masks and corresponding volumetric data. The method may furthermore include coregistering the brain masks and corresponding volumetric data to a common brain space and generating registered volumetric data. The method may in addition include performing lesion segmentation to generate a lesion mask in the common brain space. The method may moreover include applying the lesion mask data to a trained prognostic model, the prognostic model having a machine learning model configured to output a probability -based prognosis of the patient’ s brain lesion from lesion mask data, where the prognostic model has been trained with a plurality of lesion masks from a plurality of brain lesion subjects that are not the patient, and outcome data from each of the plurality of brain lesion subjects. The method may also include outputting a probability-based report of prognosis of the patient. Other embodiments of this aspect include corresponding computer systems, apparatus, and computer programs recorded on one or more computer storage devices, each configured to perform the actions of the methods.

[0007] Implementations may include one or more of the following features. The method where the classification step is performed via a classification algorithm. The method where the processing step further may include a volumetric data conversion step, where the data series of interest is converted to volumetric data of interest. The method where the processing step further may include a quality assurance step, where the volumetric data of interest may include a plurality of images and where the plurality of images is input into a quality assurance deep learning model, where the quality assurance deep learning model assigns a quality score to each of the plurality of images and where a set of images having a quality score above a predetermined threshold are selected and synthesized into one or more anatomical volumetric sequences. The method where the processing step further may include performing brain masking and skull stripping on the anatomical volumetric sequences. The method where the co-registration step further may include generating transform files from the volumetric sequence data and brain mask and resampling the volumetric data and transform files to generate co-registered volumetric data. The method where prior to the lesion segmentation step, image intensity is normalized based on intensity of the image foreground, where image intensity is normalized without clipping. The method where the patient prognosis is determined through comparison of lesion location to outcome data from other data, where the other data may include outcome data embedded in lesion-symptom maps. The method where the step of outputting a probability-based report of prognosis further may include outputting therapeutic guidance, where the therapeutic guidance is based on the prognosis output. Implementations of the described techniques may include hardware, a method or process, or a computer tangible medium.

[0008] Further disclosed herein is a system for assessing prognosis of a brain lesion patient comprising one or more processors configured to: obtain imaging data of the brain of the patient from a scanner and classify the data by acquisition plane and modality to select a data series of interest; process the data series of interest to generate brain masks and corresponding volumetric data; co-register the brain masks and corresponding volumetric data to a common brain space and generate registered volumetric data; perform lesion segmentation to generate a lesion mask in the common brain space; and apply the lesion mask data to a trained prognostic model, the prognostic model comprising a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, wherein the prognostic model has been trained with a plurality of lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects; and output a probability-based report of prognosis of the patient.

[0009] Further disclosed herein is a non-transitory computer-readable medium storing a set of instructions for assessing prognosis of a brain lesion patient, the set of instructions comprising one or more instructions that, when executed by one or more processors of a device, cause the device to: obtain imaging data of the brain of the patient from a scanner and classify the data by acquisition plane and modality to select a data series of interest; process the data series of interest to generate brain masks and corresponding volumetric data; co-register the brain masks and corresponding volumetric data to a common brain space and generate registered volumetric data; perform lesion segmentation to generate a lesion mask in the common brain space; and apply the lesion mask data to a trained prognostic model, the prognostic model comprising a machine learning model configured to output a probability -based prognosis of the patient’s brain lesion from lesion mask data, wherein the prognostic model has been trained with a plurality of lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects; and output a probability-based report of prognosis of the patient. BRIEF DESCRIPTION OF DRAWINGS

[0010] FIG. 1A is a flow diagram of the automated prediction tool, according to one implementation.

[0011] FIG. IB is a flow diagram of the automated prediction tool, according to one implementation.

[0012] FIG. 2 is a Decision Tree for Acquisition Plane Classification used in the system, according to one implementation.

[0013] FIG. 3 is a Confusion Matrix for Modality Classification used in the system, according to one implementation.

[0014] FIG. 4 Shows Axial slice of a Trace-Weighted MRI in the left posterior quadrant, transformed to MNI space.

[0015] FIG. 5 is a flow chart showing one disclosed process, according to certain embodiments.

[0016] FIG. 6 is a flow chart showing one disclosed process, according to certain embodiments.

[0017] FIG. 7 is a graph plotting the relationship between the predicted and observed scores of a disclosed process, according to certain embodiments.

DETAILED DESCRIPTION

[0018] Before explaining at least one embodiment of the invention in detail, it is to be understood that the invention is not limited in its application to the details of construction and to the arrangements of the components set forth in the following description or illustrated in the drawings. The invention is capable of other embodiments and of being practiced and carried out in various ways. Also, it is to be understood that the phraseology and terminology employed herein are for the purpose of description and should not be regarded as limiting.

[0019] Ranges can be expressed herein as from “about” one particular value, and/or to “about” another particular value. When such a range is expressed, a further aspect includes from the one particular value and/or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent “about,” it will be understood that the particular value forms a further aspect. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint. It is also understood that there arc a number of values disclosed herein, and that each value is also herein disclosed as “about” that particular value in addition to the value itself. For example, if the value “10” is disclosed, then “about 10” is also disclosed. It is also understood that each unit between two particular units is also disclosed. For example, if 10 and 15 are disclosed, then 11, 12, 13, and 14 are also disclosed.

[0020] As used herein "brain lesion patient" refers to an individual diagnosed with, or suspected to have, one or more abnormal areas of tissue within the brain that differ in structure or function from the surrounding healthy brain tissue. These abnormalities, termed "lesions," can arise from various pathological processes, including but not limited to: stroke or cerebrovascular disease; ischemic or hemorrhagic events leading to tissue death or damage in specific brain regions; tumors or neoplasms: abnormal growths, whether benign or malignant, that alter normal brain structure and function; neuroinflammatory disorders: conditions like multiple sclerosis (MS), where demyelination or inflammation causes lesions visible in neuroimaging; infections: such as brain abscesses or encephalitis, which can produce localized damage to brain tissue; congenital malformations or acquired abnormalities: including cysts or other structural anomalies visible as lesions on imaging; Traumatic Brain Injury (TBI): damage resulting from external physical force causing localized brain tissue disruption.

[0021] Some portions of this description describe the embodiments of the invention in terms of algorithms and symbolic representations of operations on information. These algorithmic descriptions and representations are commonly used by those skilled in the data processing arts to convey the substance of their work effectively to others skilled in the art. These operations, while described functionally, computationally, or logically, are understood to be implemented by computer programs or equivalent electrical circuits, microcode, or the like. Furthermore, it has also proven convenient at times to refer to these arrangements of operations as modules, without loss of generality. The described operations and their associated modules may be embodied in software, firmware, hardware, or any combinations thereof.

[0022] Any of the steps, operations, or processes described herein may be performed or implemented with one or more hardware or software modules, alone or in combination with other devices. In one embodiment, a software module is implemented with a computer program product comprising a computer-readable medium containing computer program code, which can be executed by a computer processor for performing any or all of the steps, operations, or processes described.

[0023] A method for analyzing brain imaging data such as MRI, fMRI, CT, or PET scans, from a brain lesion patient and generating prognostic information is disclosed. The method includes obtaining imaging data of the brain of a patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest. The method further involves processing the data series of interest to generate brain masks and corresponding volumetric data. In certain implementations, the brain masks and corresponding volumetric data are then coregistered to a common brain space, and registered volumetric data is generated. In certain embodiments, lesion segmentation is performed to generate a lesion mask in the common brain space. In certain alternative implementations, lesion segmentation is performed in native brain space. The lesion mask data is applied to a trained prognostic model, which comprises a machine learning model configured to output a probability-based prognosis of the patient's brain lesion from lesion mask data. The prognostic model has been trained with a plurality of lesion masks from a plurality of brain lesion subjects that are not the patient and behavioral outcome data from each of the plurality of brain lesion subjects. The method further involves outputting a probability -based report of prognosis of the patient.

[0024] In certain embodiments, the co-registration step further comprises generating transform files from the volumetric sequence data and brain mask. In certain implementations, this step includes resampling the volumetric data and transform files to generate co-registered volumetric data.

[0025] Fig. 5 is a flowchart of an example process 500. In some implementations, one or more process blocks of Fig. 5 may be performed by a system.

[0026] As shown in Fig. 5, process 500 may include obtaining imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest (block 502). For example, the system may obtain imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest, as described above. As also shown in Fig. 5, process 500 may include processing the data series of interest to generate brain masks and corresponding volumetric data (block 504). For example, the system may process the data series of interest to generate brain masks and corresponding volumetric data, as described above. As further shown in Fig. 5, process 500 may include co-registering the brain masks and corresponding volumetric data to a common brain space (c.g., Montreal Neurological Institute (MNI)) and generating registered volumetric data (block 506). For example, the system may co-register the brain masks and corresponding volumetric data to a common brain space and generating registered volumetric data, as described above. As also shown in Fig. 5, process 500 may include performing lesion segmentation to generate a lesion mask in the common brain space (block 508). For example, the system may perform lesion segmentation to generate a lesion mask in the common brain space, as described above. In alternative embodiments, lesion segmentation is performed in the patient’s native brain space.

[0027] In certain embodiments, prior to the lesion segmentation step, image intensity is normalized based on intensity of the image foreground. In exemplary implementations, image intensity is normalized without clipping.

[0028] As further shown in Fig. 5, process 500 may include applying the lesion mask data to a trained prognostic model, the prognostic model having a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, where the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects (block 510). For example, the system may apply the lesion mask data to a trained prognostic model, the prognostic model having a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, where the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects, as described above. As also shown in Fig. 5, process 500 may include outputting a probability-based report of prognosis of the patient (block 512). For example, the system may output a probability-based report of prognosis of the patient, as described above. In certain implementations, the method utilizes Lesion-symptom maps and lesion-network maps — techniques that provide crucial insights into brain-behavior relationships and the impact of focal brain damage. Lesion symptom maps statistically illustrate the association between lesion locations and specific symptoms by analyzing normalized brain images from multiple patients, highlighting areas critical for particular functions. Lesion network maps extend this concept by incorporating brain connectivity data, revealing how localized damage affects broader brain networks. These maps are created by identifying lesion locations associated with specific symptoms and using functional connectivity data to map the brain networks connected to these sites. In exemplary implementations, in this method the regression model compares the new lesion mask to the predetermined symptom and network maps to predict personalized patient outcome.

[0029] Fig. 6 is a flowchart of another example process 600. In some implementations, one or more process blocks of Fig. 6 may be performed by a system.

[0030] As shown in Fig. 6, process 600 may include obtaining imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest (block 602). For example, the system may obtain imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest, as described above. As also shown in Fig. 6, process 600 may include processing the data series of interest to generate brain masks and corresponding volumetric data (block 604). For example, the system may process the data series of interest to generate brain masks and corresponding volumetric data, as described above. As further shown in Fig. 6, process 600 may include co-registering the brain masks and corresponding volumetric data to a brain space template and generating registered volumetric data (block 606). For example, the system may coregister the brain masks and corresponding volumetric data to a brain space template and generating registered volumetric data, as described above. As also shown in Fig. 6, process 600 may include performing lesion segmentation to generate a lesion mask in the brain space template (block 608).

[0031] As further shown in Fig. 6, process 600 may include applying the lesion mask data to a trained prognostic model, the prognostic model having a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, where the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects (block 610). For example, the system may apply the lesion mask data to a trained prognostic model, the prognostic model having a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, where the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects, as described above. As also shown in Fig. 6, process 600 may include outputting a probability-based report of prognosis of the patient (block 612). For example, the system may output a probability-based report of prognosis of the patient, as described above. In certain implementations, the method utilizes Lesion-symptom maps and lesion-network maps — techniques that provide crucial insights into brain-behavior relationships and the impact of focal brain damage. Lesion symptom maps statistically illustrate the association between lesion locations and specific symptoms by analyzing normalized brain images from multiple patients, highlighting areas critical for particular functions. Lesion network maps extend this concept by incorporating brain connectivity data, revealing how localized damage affects broader brain networks. These maps are created by identifying lesion locations associated with specific symptoms and using functional connectivity data to map the brain networks connected to these sites. In exemplary implementations, in this method the regression model compares the new lesion mask to the predetermined symptom and network maps to predict personalized patient outcome.

[0032] In various embodiments, the machine learning model may be a deep learning model that may analyze additional data. As used herein, “deep learning model" refers to an artificial neural network architecture designed with multiple layers of processing units (neurons) which has been trained to recognize patterns, features, or relationships within complex neuroimaging data. This type of model leverages vast amounts of imaging data, such as MRI, fMRI, CT, or PET scans, and in certain embodiments, the deep learning model analyzes the entire multi-channel image input. Deep learning models may be comprised of multiple layers — such as input, hidden, and output layers — allowing the model to transform raw imaging data (e.g., multi-channel image input) progressively into high-level features relevant to prognosis prediction tasks. Each layer captures different levels of detail, from simple pixel-based features to complex brain patterns. In some embodiments, the additional data analyzed by the deep learning model can include a binary lesion segmentation mask, patient clinical magnetic resonance I MRI images (such as but not limited to T1 -weighted images), brain imaging data from non-lesion subjects, network atlas data, or other similar data. In certain implementations, the deep learning model generates a prognosis prediction without use of predetermined lesion symptom or network maps.

[0033] In certain embodiments, outcome data useful in training the prognostic model may include additional data such as subject demographics and other health information. These data can be entered into statistical models to improve performance of outcome predictions. In further embodiments, the additional data is optional, and prognosis can be determined using exclusively the information contained from the brain imaging.

[0034] Processes 500 and 600 may include additional implementations, such as any single implementation or any combination of implementations described below and/or in connection with one or more other processes described elsewhere herein. In a first implementation, the processing step further comprises a volumetric data conversion step, where the data series of interest is converted to volumetric data of interest.

[0035] In a second implementation, alone or in combination with the first implementation, the processing step further comprises a quality assurance step, where the volumetric data of interest comprise a plurality of images and where the plurality of images is input into a quality assurance deep learning model. In certain implementations of these embodiments, the quality assurance deep learning model assigns a quality score to each of the plurality of images. Once assigned, a set of images having a quality score above a predetermined threshold are selected and synthesized into one or more anatomical volumetric sequences. In certain implementations, the processing step further includes performing brain masking and skull stripping of the anatomical volumetric sequences.

[0036] According to certain embodiments, the patient prognosis reports are used to improve patient care in a number of ways. For instance, a stroke involving a high-risk brain region prompts more aggressive interventions in the acute stage of a stroke, such as administering thrombolytics or undergoing a clot extraction procedure. Prognostic information may also be used to guide personalized rehabilitation, such as highlighting the need for specific types of therapy or the need for additional referrals (e.g. a neuropsychological referral for an individual with a lesion location that places them at high risk for cognitive deficits. In certain embodiments, lesion location information aids in the identification of which patients may be most suitable for augmenting rehabilitation with other treatments like neuromodulation. In further embodiments, the patient prognosis report allows for more efficient and personalized inquiries on patient follow-up. For a specific example, if a patient is at high risk for post-stroke pain or depression the clinician could increase screening of those symptoms.

[0037] Although Figs. 5 and 6 show example blocks of processes 500 and 600, respectively, in some implementations, processes 500 or 600 may include additional blocks, fewer blocks, different blocks, or differently arranged blocks than those depicted in Figs. 5 and 6. Additionally, or alternatively, two or more of the blocks of processes 500 and/or 600 may be performed in parallel.

[0038] According to various embodiments, the disclosed methods are performed according to the following data flow:

1. DICOM Selection.

2. Data Processing.

3. Template Brain Registration.

4. Automated Legion Segmentation.

5. Outcome Prediction.

[0039] According to certain embodiments, the disclosed methods are performed according to the following data flow:

1. DICOM Selection

• Input: Raw DICOM data

• Output: Modality and acquisition plane classification for DICOM sequences used to select data of interest

2. Data Processing

(a) Volumetric data conversion

• Input: DICOM series of interest

• Output: Volumetric data of interest

(b) Quality Assurance

• Input: Volumetric data of interest

• Output: Quality Assurance scores used to select best quality data for further processing

(c) Image Normalization and Super-Resolution

• Input: Anatomical Volumetric sequence with the best quality score

• Output: Synthesized high-resolution Tlw image

(d) Brainmasking and Skull-Stripping

• Input: Volumetric data of interest

• Output: Brainmasks and skull- stripped volumes

3. MNI152 Registration (a) Co-registration

• Input: Volumetric sequences and brainmasks

• Output: Transform files

(b) Resampling

• Input: Volumetric data and transforms files

• Output: Co-registered volumetric data

(c) MNI Registration

• Input: Original/Synthesized Tlw sequence

• Output: Native image space to MNI152 Transform

(d) Resampling to MNI

• Input: Co-registered Volumetric data and MNI Transform

• Output: MNI registered data.

4. Automated Lesion Segmentation

• Input: Volumetric data sequences in MNI space

• Output: Lesion Mask in MNI space

5. Outcome Prediction

• Input: Volumetric data and Lesion Mask

• Output: Predicted Outcome Scores, which could include any functional outcome related to lesion location, such as cognition, motor outcomes, disability, mood, personality, pain, or others.

[0040] While legion segmentation described above is performed in the MNI space, this is merely exemplary of one embodiment. Other spatial systems, such as but not limited to common brain space or native brain space, can be used.

[0041] While multiple embodiments are disclosed, still other embodiments of the disclosure will become apparent to those skilled in the art from the following detailed description, which shows and describes illustrative embodiments of the disclosed compositions, systems and methods. As will be realized, the disclosed compositions, systems and methods are capable of modifications in various obvious aspects, all without departing from the spirit and scope of the disclosure. Accordingly, the drawings and detailed description are to be regarded as illustrative in nature and not restrictive. EXAMPLES

[0042] The following examples arc put forth so as to provide those of ordinary skill in the art with a complete disclosure and description of certain examples of how the compounds, compositions, articles, devices and/or methods claimed herein are made and evaluated, and are intended to be purely exemplary of the invention and are not intended to limit the scope of what the inventors regard as their invention. However, those of skill in the art should, in light of the present disclosure, appreciate that many changes can be made in the specific embodiments which are disclosed and still obtain a like or similar result without departing from the spirit and scope of the invention.

Example 1

Overview

[0043] In this example, an automated tool was developed for predicting stroke outcomes in acute ischemic stroke patients. As shown in Fig. 1A, the tool used in this example has five main steps: Digital Imaging and Communications in Medicine (“DICOM”) Selection (with several substeps), Data Preprocessing, MNI152 Registration, Lesion Segmentation, and Lesion Segmentation Outcome Prediction. The output of this tool is reliable prognosis for a particular patient based on their CT or MRI scan data.

Data Selection

[0044] To ensure consistency and adherence to contemporary protocols in prospective studies, the processing pipeline mandated specific MRI sequences. These sequences included an anatomical scan (either T1 -weighted, T2- weighted, or Fluid- Attenuated Inversion Recovery (FLAIR)), Apparent Diffusion Coefficient (ADC) and Trace- Weighted images. This stringent set of imaging requirements is vital for the robust analysis conducted by the tool.

[0045] The initial cohort had 138 subjects. Upon applying the inclusion criteria, 119 were confirmed to have acute ischemic strokes. Notably, the original retrospective dataset had been processed manually, with available sequences determining subject inclusion. Aligning with current standards for prospective studies, the pipeline necessitated the complete suite of specified modalities for analysis. Consequently, 5 subjects were excluded due to the absence of the requisite MRI sequence data, resulting in a refined final cohort of 114 patients. This exclusion criterion is a deviation from past manual processes and is reflective of a commitment to the integrity of automated, protocol-driven research. All patients included were adults over 18 years of age, with no history of alcohol or substance abuse, psychiatric disorders, or neurological conditions unrelated to the ischemic lesion. They had undergone neuropsychological testing at the University of Iowa Benton Neuropsychology Clinic, following the protocols of the Benton Neuropsychology Laboratory. The cognitive function of participants was evaluated using a comprehensive battery of neuropsychological tests.

DICQM Selection

[0046] For this example, a robust, easily extensible classification framework that extracts key features from well-characterized DICOM header fields to identify image modality and acquisition plane with over 99% was developed. The dcm.classifier tool allows the selection of data volume needed for completing the analysis, thus eliminating a bottleneck of manual selection of the DICOM objects. The tool is described in the following steps in detail.

DICOM Selection Step 1

[0047] A training data set is required for the Example’s classification models. PyDICOM was used to extract the DICOM header information for every scan in the dataset. Certain specific fields are especially prone to human errors. These fields were identified and were used to infer the type of scan, such as CT scan or MRI scan, and its orientation. From this classification, eleven features were extracted. The features included a combination of raw DICOM fields and their derivatives. The raw DICOM fields included:

• (0018,0081) EchoTime - Time in ms between the middle of the excitation pulse and the peak of the echo produced;

• (0018,0080) RepetitionTime - The period of time in ms between the beginning of a pulse sequence and the beginning of the succeeding pulse sequence;

• (0018,1314) FlipAngle - Steady state angle in degrees to which the magnetic vector is flipped from the magnetic vector of the primary field;

• (0018,0095) PixelBandwidth - Reciprocal of the total sampling period, in hertz per pixel; and

• (0018,1316) SAR - Calculated whole body Specific Absorption Rate in watts per kilogram. [0048] Six more features were engineered to provide the model with sufficient context. A count of the number of repeated (0020,0032) Image Position (Patient) values represents the number of series’ volumes. Next, based on (0008,0008) ImageType - array containing image identification characteristics, three binary flags were created for TraceW, ADC, and FA images. Finally, b-value information was extracted to identify TraceW Diffusion-Weighted Imaging (DWI) scans.3 This process is manufacturer dependent and uses field (0018, 9087) for standard, (0043, 1039) for GE, (2001, 1003) for Philips, (0019, 100C) for Siemens ((0019, 000C) for old Siemens data), or (0065, 1009) for UHI data. All distinct b- values were combined into an array and create two fields: maximal b-value and a number of distinctive b-values. Separately, for image acquisition plane classification, (0020,0037) ImageOrientation (Patient) field was converted from an array into 6 separate features (of which only 2 are sufficient for model prediction).

DICOM Selection Step 2- Acquisition Plane Classification Model

[0049] The acquisition plane was automatically classified using a Decision Tree algorithm. At each split, the algorithm maximizes the reduction in impurity or variance, resulting in a treelike structure where each node represents a decision based on a feature value. The final predictions are made by traversing the tree from the root node to a leaf node corresponding to the specific class.

[0050] Using six components of (0020,0037) ImageOrientation (Patient) DICOM field as input, a Decision Tree was trained with a maximum depth of two. Analyzing the produced model revealed that only two of six features were meaningful to the classification and multiple combinations of those features can achieve accuracy of 99.5% or above.

[0051] The model performed at 99.96% accuracy on the test dataset. On 1453 axial, 303 coronal, and 510 sagittal scans in the test dataset, the model achieved precision and recall of 1.00 for each class and misclassified only a single sagittal image as coronal. It is important to note however, that imaging is rarely done in the perfectly square direction system, and at 45 degrees classification is ambiguous.

[0052] The model converged to the values of 0.56 and 0.44 for the sixth and second component of the (0020,0037) ImageOrientation (Patient) DICOM field, as shown in Fig. 2. Theoretically, those values should be 0.5. However, imaging is rarely done in the perfectly square direction system, and the values were optimized based on specific datasets. DICOM Selection Step 3 - Modality Classification Model

[0053] The same image types are often collected with identical MRI device settings and protocols. This is especially common in datasets from the same site. After the DICOM Feature Extraction step (Experimental Step 1), duplicates were found. To properly investigate this issue, all features were rounded to three significant digits based on the maximal feature value (if maximal (0018,0080) RepetitionTime = 9000, then 4327 4330 — > 4.52 becomes 0). This resulted in 3139 unique datasets and ensured the model training and evaluation were not using the same data.

[0054] The Random Forest algorithm was selected for the model in this Experimental Step. The Random Forrest Classifier using Scikit-Leam was implemented, and multiple model configurations were tested. The finally selected configuration used Random Forrest with 100 estimators, a maximum depth of 7, and a Gini impurity criterion. The model input is the array of 11 features described in Experimental Step 1, and the output is the probability for each 11 modality classes, where the highest probability is the model prediction.

[0055] During the model hyperparameter tuning, multiple maximum depth settings were used. A high 98.8% accuracy for the reported model was achieved, which was further confirmed by performing K-Fold Cross Validation. It was found that scaling the model past the max depth of 7 did not affect the accuracy on the Brain and Prostate dataset but started overfitting on the Stroke dataset, which contains known mistakes.

[0056] A confusion matrix was generated for modality classification on the test dataset to get insight into the model behavior, as seen in Fig. 3. It was observed that the model accurately predicted all images for most modalities. Most mistakes involved misclassifying FLAIR images as T2w images. This can be explained by the close nature of those modalities.

[0057] Additionally, the top Decision Tree estimators created during the random forest model training were inspected. A strong correlation to MRI parameters that could be found in the literature was found through this inspection. The EchoTime and RepetitionTime features had thresholds especially similar to values of the typical MRI protocols.

DICOM Selection Step 4a - K-Folcl Cross-Validation

[0058] A K-Fold Cross-Validation analysis provides better insights into the model behavior and confirms the robustness of the high-accuracy results. This technique reduces the risk of overfitting by splitting the available data into equally sized K subsets, then training and evaluating the model K times. In each of the K iterations, one fold is used as the test set, and the other folds are used as the training set. In this validation, K=5 was used to get a standard 80-20 % split for training and validation. The folds are taken from the unique dataset and matched with the original data. This ensures that the same data does not appear in the training and test datasets at any time during the k-fold cross-validation.

[0059] The model achieved a total accuracy of 99.4% with a standard deviation of 0.5% for the K=F-Fold Cross- Validation. For seven modalities, the model achieved an average score of 1.00 with a 0.00 standard deviation for precision, recall, and f-1 score. Table 1 displays the averaged results for the remaining four modalities.

Table 1. Averaged results of K-Fold Cross-Validation for modality classification. For classes without perfect scores, the mean and the standard deviation of precision, recall, and f-1 score are presented. The other 7 modalities not shown had perfect precision, recall and Fl -score.

DICOM Selection Step 4b - Independent Dataset Validation

[0060] To further validate the experimental model, the model was tested on an independent dataset of Minipig Head MRI. This dataset contained 357 datasets across four modalities and was collected using a body coil rather than the typical bird-cage coil. Table 2 displays precision, recall and fl -score metrics and the number of images of each modality in the support column. For each of the dataset modalities, a very high precision and recall was observed, except for the Field Map, which had a much lower recall at 0.79. Overall, the model performed at an accuracy of 98.6% - comparable to the performance on human data. Interestingly, however, the model misclassified one DWI gradient and Field Map image as TraceW, although this data type was absent in the dataset.

Table 2. Modality classification performance on out-of-sample minipig dataset. The model achieved performance similar to the human data at 98.6% accuracy. The model misclassified one dataset as TraceW - modality not present in the dataset.

Data Preprocessing

[0061] The data processing workflow has four main steps, each tailored to enhance the quality and usability of MRI data for stroke outcome analysis. First, the DICOM images from the chosen modalities — T1 -weighted, T2- weighted, FLAIR, ADC, and Trace-Weighted — are transformed into NlfTI format via the dcm2niix tool and ITK. This step facilitates the transition from individual 2D slices to integrated 3D volumes, essential for subsequent spatial analyses.

[0062] Second, due to the inherent variability in scan quality — often influenced by factors such as artifacts or patient movement — it becomes necessary to ensure the selection of high-quality data. To this end, the MIQA quality assurance tool was employed, which leverages a deep learning model specifically trained for image quality assessment. This allows the objective selection of the highest-quality images for each modality. When faced with multiple sequences of the same modality, preference is given to the sequence that the MIQA tool rates as the highest quality.

[0063] The third step involves taking the best-quality anatomical scan and synthesizing a high-resolution, normalized Tl- weighted image suitable for registration. For this purpose, the SynthSR tool was used. The final step is the performance of brain masking and skull stripping using SynthStrip. The brainmasks are critical for defining the specific region for registration, while the skull-stripped data is a prerequisite for the accurate segmentation of brain lesions.

MNI152 registration

[0064] Co-registration of MRI images is a pivotal step in the workflow due to the inherent distortions that may arise even within the same scanning session. Although images are captured in close temporal proximity, different modalities can act like varied lenses, each introducing its own distortions. Additionally, involuntary patient movement during the session can lead to misalignment, which can negatively impact further processing. Thus, precise image registration is imperative.

[0065] Co-registration was executed by selecting the highest quality anatomical scan (Tlw, T2w, or FLAIR) as the registration target. Subsequently, the ADC and Trace-Weighted images are co-registered to the target using rigid and affine transformations, aligning all images within the same physical space. This step is enhanced by the utilization of the brainmask for each image, which facilitates more precise sampling during co-registration.

[0066] The second phase involves transforming the images to the MNI space. This standard space is widely recognized in neuroimaging and is particularly indispensable for lesionbehavior mapping studies, which necessitate that all data be situated within a uniform physical reference frame.

[0067] Challenges in this transformation process emerge with lesions, which represent areas of brain damage and can affect the accuracy of registration to templates typically derived from healthy brains. Nonetheless, acute ischemic stroke lesions often do not manifest distinctly on the anatomical scans used for registration. Moreover, the high-resolution Tlw image generated by the previous step effectively normalizes any pathological changes.

[0068] To execute the MNI152 transformation, a rigid, affine, and SyN (non-linear) transformations using the BRAINSFit program from the BRAINSTools package was employed. Upon determining all transformations, the data was resampled to conform to the MNI152 Tlw template space.

[0069] The data resulting from registration and all other processes withing the tool adheres to the Brain Imaging Data Structure (BIDS) standard, which ensures systematic and accessible organization of both raw and derived datasets.

Lesion Segmentation

[0070] The accurate segmentation of focal lesions is integral, as outcome-predictive methodologies are contingent upon precise lesion identification. Recent advancements in computational methodologies have positioned deep learning as the vanguard approach for lesion segmentation. [0071] The refined model capitalizes on the diagnostic strengths of Trace-Weighted and ADC images. It was rigorously trained on a dataset comprising approximately 344 retrospective cases of acute ischemic stroke, sourced from the University of Iowa Hospital and Clinics. A key metric of its performance — the DICE coefficient — stands at a robust 0.74, signifying a high degree of accuracy in lesion delineation considering the original data resolution. It is important to mention that the dataset utilized in the current study was not previously encountered by this model, thereby providing an unbiased test of its generalizability. An axial slice of a Trace-Weighted MRI in the left posterior quadrant, transformed into MNI space. Red arrows indicate the lesion’s extent as delineated by the automated segmentation model (green outline) and a brain region critical for cognition (highlighted in red). This patient’s imaging demonstrates a substantial overlap of the lesion with the region of interest, suggesting this individual is at high risk for chronic impairments in cognition.

Lesion-Based Outcome Prediction

[0072] This example leverages the predictive power of specific brain regions within the left posterior white matter concerning general cognitive impairment. The employed methods and tools have automated the categorization of subjects into high and low-risk groups, based on the overlap of the lesion masks — generated in the segmentation step — with a priori defined regions of interest (“ROI”s).

[0073] The initial ROI was derived from the most statistically significant anatomical cluster on the multivariate lesion- symptom map of general cognitive ability (g). This cluster was situated at the intersection of several white matter association tracts, notably featuring the arcuate fasciculus. Additionally, an ROI derived from an edge density map was incorporated, employing a threshold that has been previously shown to correlate with significant cognitive deficits.

[0074] The proportion of impaired cognitive tests served as the primary outcome measure, comparing the high and low-risk groups within each ROI. The Shapiro-Wilk test was utilized to assess the normality of the data distribution. Given the nonnormal nature of the data, the Mann- Whitney U-test was applied for statistical comparison. The analysis was further refined through ANCOVA to control for variables that could potentially confound the findings, specifically age, time since lesion onset, and lesion volume. Experimental Results

[0075] Within the first region of interest (ROI), high risk patients exhibited a mean Z-scorc of -6.00 and an average proportion of impaired tests of 0.73±0.31. This contrasts with low-risk patients, who demonstrated a mean Z-score of -1.51 and an average proportion of 0.33 ± 0.26. The Mann- Whitney U-test yielded a p-value of less than 0.0001, indicating a highly significant statistical association between lesions involving the ‘g’ region and post-stroke general cognitive impairment. Furthermore, these results remained significant in an Analysis of Covariance (ANCOVA) that accounted for lesion volume, patient age, and time elapsed since stroke onset, with a p-value of 0.000009.

[0076] In the analysis of the second ROI, which focused on the edge density map, high- risk patients had a mean Z-score of -3.80 and an average proportion of impaired tests of 0.52 ± 0.33, whereas low-risk patients had a mean Z-score of -1.34 and an average proportion of 0.32±0.25. The corresponding p-values from the Mann -Whitney U-test and the ANCOVA were 0.003 and 0.049, respectively, suggesting statistical significance.

[0077] A pivotal aspect of the tool’s design is its fully automated operation, which necessitates only two inputs: the directory path to the patient’s raw DICOM data directory and the destination path for the Brain Imaging Data Structure (BIDS) formatted results. On average, the pipeline reaches completion in approximately 12 minutes when executed on a standard CPU. It is noteworthy that a substantial portion of this duration — up to 10 minutes — is dedicated to the computationally intensive task of the non-linear registration step.

Example 2

Overview

Artificial Intelligence (Al) has emerged as a transformative force in healthcare. By combining software engineering, medical image processing and Al techniques researchers created many complex processing pipelines in a variety of applications from neuroimaging, cancer detection and even animal studies. Such techniques, bring possibility of automated processing suitable for clinical integration, reduction of error-prone human intervention and ability to process and leverage vast amounts of data resulting in significant acceleration of research and clinical adaptation. [0078] This example presents a fully-automated system, diagrammed in Fig. IB, that integrates advanced image processing techniques with machine learning algorithms to generate personalized stroke outcome predictions. The system processes patient data and delivers individualized prognoses within minutes, facilitating evidence-based decision-making in timesensitive scenarios. The automated pipeline represents a significant step towards the clinical implementation of systems that can enhance the planning and personalization of rehabilitation strategies for stroke survivors. The technology disclosed in this example has the potential to significantly advance the standard of care and improve recovery outcomes for stroke patients.

Methods

DICOM Data Selection

[0079] The analysis process can begin with extracting key features from well-characterized DICOM header fields to identify image modality and an acquisition plane with high accuracy, such as over 99% accuracy. The capabilities of the system in this example were expanded with post-processing functionality that allows to automatically select the modalities of interest and transform them into volumetric representation utilizing Neuroimaging Informatics Technology Initiative (NIITI) or Nearly Raw Raster Data (NRRD) format. The selection of data can be motivated by standard post-stroke imaging protocols and can include: Tl-weighted, T2-weighted, Fluid-Attenuated Inverse Recovery (FLAIR), Apparent Diffusivity Coefficient (ADC), and the Diffusion Trace-Weighted acquired in axial direction or isotropic. This fully-automated procedure ensures the robust selection of input data that is scalable in a clinical environment.

Data Preprocessing

[0080] In the embodiment described in this example, the data processing workflow has three main steps, each tailored to enhance the quality and usability of MRI data for stroke outcome analysis. First, the volumetric data created by the previous step can be verified. The presence of ADC and Trace-Weighted and one of the structural images (Tlw, T2w or FLAIR) is required for system functionality. Second, data quality assurance can be performed using the Medical Image Quality Assurance (MIQA) quality assurance tool, which leverages a deep learning model specifically trained for image quality assessment. This allows the objective selection of the highest-quality images for each modality and best structural image overall, which can serve as the target image for subsequent registration.

[0081] Third, data derivatives can be created as follows. From the target image, a high- resolution, normalized T 1 -weighted image can be created using SynthSR. For each image, a binary brain mask was created using a custom trained model based on the ResUNet architecture, which achieved > 0.98 DICE score compared to ground-truth segmentations. The brain masks are also used to create the skull-stripped images which will be utilized in further parts of the system.

[0082] Automated quality control mechanisms are integrated throughout these steps to identify and eliminate any processing failures, ensuring that only data of the highest integrity progresses to subsequent analyses. This data processing paradigm consistently achieves a processing time of less than one minute, with a success rate exceeding 99% in either successfully processing the data or accurately flagging issues that could lead to downstream errors. The combination of high efficiency and robust performance demonstrates that this approach is well- suited for seamless integration into clinical workflows.

Template Brain Registration

[0083] Despite the temporal proximity of the image acquisition within a single scanning session, different modalities and potential patient movement can introduce misalignments, necessitating precise image registration to ensure accurate subsequent processing, the subsequent task of the lesion-based outcome prediction utilizes lesion-symptom mapping techniques are defined in the Montreal Neurological Institute (MNI) space, a widely recognized standard in neuroimaging, and thus requiring us to transform patient data to the MNI space.

[0084] First, the highest quality images are co-registered per modality to the target image (as defined in DICOM Data Selection). Second, the Tl-weighted image, which was synthesized from the target image, is used to create the transform from the patient native space to MNI space. The co-registered data can then be leveraged to resample all images to the template space. Both co-registration and MNI registration use custom-implemented with ITK Rigid and Affine transformations. MNI registration also performs non-linear SyN transformation utilizing Advanced Normalization Toolkit (ANTs). All stages utilize the pre-computed brain masks to define the registration sampling region. [0085] While registering to the healthy /normal brain in template space, presence of lesions can affect registration accuracy. However, in the acute stage, ischemic stroke lesions often do not manifest on structural images. Furthermore, the SynthSR tool used to synthesize Tlw image in preprocessing step, effectively normalizes pathological changes and improves registration.

[0086] We tested this system on over 3000 individual images. The results were both visually inspected and confirmed by metric defined as DICE score of over 0.95 between the registered brain mask and MNI template brain mask. The automated registration framework of this example achieved > 99% in successful co-registration and MNI registration demonstrating reliable performance suitable for clinical integration.

Lesion Segmentation

[0087] This example produces a clinical tool that can automatically leverage clinical imaging data to generate individual predictions of cognitive outcomes after stroke. To demonstrate the feasibility of this goal, a lesion-based outcome prediction model for the Boston Naming Test (BNT) was developed and incorporated this prediction model into our automated pipeline. The BNT is a commonly used assessment of post-stroke language function. It has previously been shown that post-stroke BNT performance can be successfully predicted in independent datasets using information about lesion location (i.e., lesion-behavior maps) and its relationship to distributed structural and functional brain networks (i.e., lesion-network maps) making it well- suited for demonstrating the feasibility of automated cognitive outcome prediction. Lesionbehavior and lesion-network maps were used to generate a simple proof-of-concept outcome prediction model that can be applied to the automatically generated lesion segmentations to obtain predicted BNT scores for individual patients as part of our end-to-end pipeline.

[0088] Model training was performed using functions implemented in the MATLAB r2022b Statistics and Machine Learning Toolbox. A ridge regression model was trained with stratified 5-fold cross-validation to tune the lambda hyperparameter. Training set patients were stratified into 4 groups based on the quartiles of the training set BNT score distribution prior to defining the train/test partitions for hyper-parameter optimization using the function cvpartition() with the “Stratify” option set to true. Model training was conducted using the function fitrlinear(). The outcome variable for the model was the normalized BNT score. The predictor data were defined as lesion loads on each of the lesion-behavior, structural lesion-network, and functional lesion-network maps, where lesion loads for each map are computed as the sum of voxel weights within each patient’s lesion, divided by the sum of all voxel values within the map. This produced 1 lesion-load predictor for the lesion-behavior map, 3 lesion load predictors for the different structural lesion-network maps, and 9 lesion load predictors for the different functional lesionnetwork maps, for a total of 13 lesion load predictors. Predictor and outcome data were standardized (i.e., converted to z-scores) prior to model training, and the center and scale parameters obtained from the training dataset were saved so that they could be applied to the test dataset prior to obtaining model predictions. The lesion load predictor computation step was embedded along with the fully trained prediction model into the automated pipeline, along with the center and scale factors obtained from the training dataset. The full end-to-end pipeline was then run to produce lesion segmentations, compute lesion load predictors, and obtain predicted BNT scores for each patient in the test dataset. The prediction performance in the test dataset was evaluated by computing the prediction R-squared (i.e., explained variance) using the sum-of- squares formulation along with the Pearson correlation between predicted and observed scores.

Data

Outcome Prediction Model Training

[0089] The outcome prediction model was trained using a sample of N=388 patients from the Iowa Lesion Registry. These patients were a subset of the N=432 sample, with 44 patients being removed from the training dataset due to also having their acute data included in the test dataset.

Clinical cohort used for out-of-sample testing

[0090] The test dataset for this example comprised retrospective clinical data from 263 acute ischemic stroke patients treated at the University of Iowa Healthcare system. All imaging data were acquired within one week of stroke onset, with sampling dates ranging from 2001 to 2024. The dataset reflects the diversity of clinical settings, encompassing images from 27 unique scanner models across 6 manufacturers. This variability in imaging sources underscores the robustness of our system across different hardware configurations.

[0091] Subjects were deemed eligible for processing if their scanning session included an ADC sequence, a Trace-Weighted image, and at least one structural image (T 1 -weighted, T2- weighted, or FLAIR). These eligibility criteria ensure the availability of essential data for our analytical pipeline while accommodating the variability in clinical imaging protocols. All eligible subjects were subsequently enrolled in a research program that included a battery of behavioral tests administered during the chronic stage of stroke. Due to various clinical and logistical factors, not all patients completed the same set of assessments. Table 3 summarizes the behavioral tests employed in this study, detailing the number of patients with available scores for each test and the number of patients classified as impaired.

Results

Outcome Prediction Results

[0092] A ridge regression model was trained on a sample of N=388 patients with chronic brain lesions from the Iowa Lesion Registry. We applied this model to predict chronic BNT scores based on lesion load variables that were automatically computed using previously published lesion-behavior and lesion-network maps, along with lesion segmentations that were automatically generated from acute diffusion weighted imaging in an independent sample of stroke patients. Of the 230 stroke patients in the test dataset, 121 had chronic timepoint data for the BNT and were included in the analysis. Comparison of the predicted and observed BNT scores in the test dataset indicated that the model successfully explained —1/3 of the total variance in BNT scores for the test dataset, with a highly significant correlation between the predicted and observed BNT scores (prediction R2=0.33, Pearson r=0.60, p<0.001).

[0093] The relationship between the predicted and observed scores is shown in Fig. 7. While the model explained a large chunk of the total variance in the test dataset (-33%), Fig. 7 shows that it often under-estimated the severity of the predicted deficits (i.e., many of the cases fall to the left of the identity line). This result demonstrates the feasibility of cognitive outcome prediction within the context of a fully automated lesion segmentation and outcome prediction pipeline.

System Performance

[0094] The system's performance was evaluated on the test dataset - retrospective clinical dataset of acute stroke patients. The system demonstrated 100% reliability, successfully processing all subjects that met the eligibility criteria (presence of ADC, Trace-Weighted, and at least one structural image). For subjects not meeting these criteria, the system reported the issue and exited gracefully, ensuring robust handling of diverse clinical data.

[0095] The system was executed entirely on a single machine with Intel Xeon CPU, including deep learning components. Most processes utilized a single core, except for the registration step, which leveraged 24 cores. Performance metrics for each system component are summarized in Table 3.

[0096] Registration accounted for the majority of the processing time, reflecting its computationally intensive nature as an iterative convergence process. This step also exhibited the highest variance in runtime, as evidenced by its standard deviation of 80 seconds. Despite this variability, the system achieved an average end-to-end processing time of 267.8 seconds (approximately 4.5 minutes), with 95% of cases completing in under 5 minutes. Such efficiency underscores the system's potential for seamless integration into time-sensitive clinical workflows, particularly in acute stroke care settings where rapid analysis is crucial.

Table 3. Runtime summary of processing system components in seconds.

Claims

CLAIMS What is claimed is:

1. A method of assessing prognosis of a brain lesion patient, the method comprising: obtaining imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest; processing the data series of interest to generate brain masks and corresponding volumetric data; co-registering the brain masks and corresponding volumetric data to a common brain space and generating registered volumetric data; performing lesion segmentation to generate a lesion mask in the common brain space; and applying the lesion mask data to a trained prognostic model, the prognostic model comprising a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, wherein the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects; and outputting a probability-based report of prognosis of the patient.

2. The method of claim 1, wherein the classification step is performed via a classification algorithm.

3. The method of claim 1, wherein the processing step further comprises a volumetric data conversion step, wherein the data series of interest is converted to volumetric data of interest.

4. The method of claim 3, wherein the processing step further comprises a quality assurance step, wherein the volumetric data of interest comprise a plurality of images and wherein the plurality of images is input into a quality assurance deep learning model, wherein the quality assurance deep learning model assigns a quality score to each of the plurality of images and wherein a set of images having a quality score above a predetermined threshold are selected and synthesized into one or more anatomical volumetric sequences.

5. The method of claim 4, wherein the processing step further comprises performing brain masking and skull stripping the anatomical volumetric sequences.

6. The method of claim 1, wherein the co-registration step further comprises generating transform files from the volumetric sequence data and brain mask and resampling the volumetric data and transform files to generate co-registered volumetric data.

7. The method of claim 1, wherein prior to the lesion segmentation step, image intensity is normalized based on intensity of the image foreground, wherein image intensity is normalized without clipping.

8. The method of claim 1, wherein the patient prognosis is determined through comparison of lesion location to outcome data from other data, wherein the other data comprises outcome data embedded in a lesion- symptom maps.

9. The method of claim 1, wherein the step of outputting a probability -based report of prognosis further comprises outputting therapeutic guidance, wherein the therapeutic guidance is based on the prognosis output.

10. A system for assessing prognosis of a brain lesion patient comprising: one or more processors configured to: obtain imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest; process the data series of interest to generate brain masks and corresponding volumetric data; co-register the brain masks and corresponding volumetric data to a common brain space and generating registered volumetric data; perform lesion segmentation to generate a lesion mask in the common brain space; and apply the lesion mask data to a trained prognostic model, the prognostic model comprising a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, wherein the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects; and output a probability-based report of prognosis of the patient.

11. The system of claim 10, wherein the one or more processors, when the processing step further, are configured to a volumetric data conversion step, wherein the data series of interest is converted to volumetric data of interest.

12. The system of claim 11, wherein the one or more processors, when the processing step further, are configured to a quality assurance step, wherein the volumetric data of interest comprise.

13. The system of claim 12, wherein the one or more processors, when the processing step further, are configured to perform brain masking and skull stripping the anatomical volumetric sequences.

14. The system of claim 10, wherein the co-registration step further comprises generating transform files from the volumetric sequence data and brain mask and resampling the volumetric data and transform files to generate co-registered volumetric data.

15. The system of claim 10, wherein prior to the lesion segmentation step, image intensity is normalized based on intensity of the image foreground, wherein image intensity is normalized without clipping.

16. A non-transitory computer-readable medium storing a set of instructions for assessing prognosis of a brain lesion patient, the set of instructions comprising; one or more instructions that, when executed by one or more processors of a device, cause the device to: obtain imaging data of the brain of the patient from a scanner and classifying the data by acquisition plane and modality to select a data series of interest; process the data series of interest to generate brain masks and corresponding volumetric data; co-register the brain masks and corresponding volumetric data to a common brain space and generating registered volumetric data; perform lesion segmentation to generate a lesion mask in the common brain space; and apply the lesion mask data to a trained prognostic model, the prognostic model comprising a machine learning model configured to output a probability-based prognosis of the patient’s brain lesion from lesion mask data, wherein the prognostic model has been trained with a plurality lesion masks from a plurality of brain lesion subjects that are not the patient and outcome data from each of the plurality of brain lesion subjects; and output a probability-based report of prognosis of the patient.

17. The non-transitory computer-readable medium of claim 16, wherein the classification step is performed via a classification algorithm.

18. The non-transitory computer-readable medium of claim 16, wherein the one or more instructions that cause the device to the processing step, cause the device to a volumetric data conversion step, wherein the data series of interest is converted to volumetric data of interest.

19. The non-transitory computer-readable medium of claim 18, wherein the one or more instructions that cause the device to the processing step, cause the device to a quality assurance step, wherein the volumetric data of interest comprise.

20. The non-transitory computer-readable medium of claim 19, wherein the one or more instructions that cause the device to the processing step, cause the device to perform brain masking and skull stripping the anatomical volumetric sequences.