Glass brains combined with felzenswalb segmentation for statistical parametric mapping optimizations

Abstract

Systems, devices, and methods for identifying and segmenting significantly different brain clusters for patients compared to normal controls for statistical parametric mapping (SPM) optimizations are provided. In one embodiment, a non-transitory machine-readable medium storing a program for SPM optimization is provided, the program comprising instructions, which when executed cause a client device to: select at least one set of t-contrast files from a patient's contrast files; generate at least one Glass Brain image using the at least one set of t-contrast files, wherein the at least one Glass Brain image includes a three-dimensional (3D) visualization of the patient's brain; perform a segmentation process on the at least one Glass Brain image, wherein the segmentation process generates at least one segmented image; generate cluster data, wherein the cluster data includes data for at least one activated cluster; and diagnose a brain condition for the patient using the cluster data.

Description

GLASS BRAINS COMBINED WITH FELZENSWALB SEGMENTATION FOR STATISTICALPARAMETRIC MAPPING OPTIMIZATIONSCROSS-REFERENCE TO RELATED APPLICATION

[0001] The current application claims priority to U.S. Provisional Patent Application No. 63 / 540,462 filed on September 26, 2023, the disclosure of which is incorporated herein by reference.FIELD OF THE INVENTION

[0002] The present invention generally relates to neuroimaging and more specifically to systems, devices, and methods for identifying and segmenting significantly different brain clusters for patients compared to normal controls for statistical parametric mapping (SPM) optimizations.BACKGROUND

[0003] Neuroimaging may include the use of quantitative and computational techniques to study the structure and function of a person’s central nervous system including the brain and the spinal cord. For example, neuroimaging may be utilized as an objective way of studying a human brain in a non-invasive manner. Neuroimaging may also be used for quantitative studies of brain disease and psychiatric illness. Functional brain imaging techniques, such as functional magnetic resonance imaging (fMRI), may also be common in neuroimaging.

[0004] Generally, neuroimaging may be categorized in two broad categories. A first category may be structural imaging, which is used to quantify brain structure using, for example, voxel based morphometry. A second category may be functional imaging, which is used to study brain function, often using fMRI and other techniques such as positron emission tomography (PET) and magnetoencephalography (MEG).SUMMARY OF THE INVENTION

[0005] The various embodiments of the present systems, methods, and devices for SPM optimizations contain several features, no single one of which is solely responsible for their desirable attributes. Without limiting the scope of the present embodiments, their more prominentfeatures will now be discussed below. In particular, the present SPM optimizations will be discussed in the context of Glass Brains and Felzenswalb segmentation. Glass Brains may include three-dimensional (3D) visualization of a patient’s brain including, but not limited to, brain differences in patients compared to normal controls. Glass Brains may be used to further understand the brain including brain-related diseases. The use of Glass Brains is merely exemplary and various other representation and / or visualization of neuroimaging may be utilized for SPM optimizations as appropriate to the requirements of a specific application in accordance with embodiments of the invention. Likewise, the use of Felzenswalb segmentation is merely exemplary and various other types of segmentation may be utilized as appropriate to the requirements of a specific application in accordance with embodiments of the invention. In addition, the present SPM optimizations will be discussed in the context of particular software such as, but not limited to, coding languages (e.g., Python) and applications (e.g., Matlab). However, the use of any particular coding language and / or application are merely exemplary and various other software may be utilized for SPM optimizations as appropriate to the requirements of a specific application in accordance with embodiments of the invention.

[0006] Further, the present SPM optimizations will also be discussed using an illustrative example. Specifically, the present embodiments will be discussed in the context of validation of [18F]-fluorodeoxyglucose (FDG) positron emission tomography (PET) and diffusion tensor imaging for diagnosis of traumatic brain injury (TBI). However, the illustrative example is merely exemplary and various other applications and / or uses may be conducted using SPM optimizations as appropriate to the requirements of a specific application in accordance with embodiments of the invention. After considering this discussion, and particularly after reading the section entitled “Detailed Description,” one will understand how the features of the present embodiments provide the advantages described here.

[0007] One aspect of the present embodiments includes the realization that as part of projects to validate parameters for SPM, area under the receiver operating characteristic curve (AUROC) analysis may be performed on a set of control and patient files (may also be referred to as “control and patient contrast files”). The control contrast files and the patient contrast files may include control contrast images and patient contrast images, respectively. During the process, leave one out (LOO) analysis may be performed to ensure there is no bias when control files are testedalongside patient (the experimental) files. However, during testing, there may be many controls registering as false positive results, when all of the controls should in fact be registering as true negatives, since none of the control files are diseased.

[0008] Another aspect of the present embodiments includes the realization that the above issue may be specific to a project currently being conducted, but there may be many solutions available to similar issues. Current solutions may include a voxel-based clustering approach, where the control files are fed into a batch-learning process, followed by K-means clustering calculations, and then the calculation of clustering maps. This approach may filter out false positives by first classifying and defining the regions of interest in magnetic resonance imaging (MRI) patient scans first, then comparing them to the control files using a Support Vector Machine (SVM). However, this method may be inefficient because support vector machine architecture utilization typically requires manually defined regions of interest (ROIs) instead of computer defined ROIs.

[0009] Another aspect of the present embodiments includes the realization that another common solution that uses receiver operating characteristic (ROC) analysis may use EZ-MAP or “one vs. many” t-test statistical analysis during the generation of the area under the curve (AUC) values. This approach discriminates the patients from the control files, counting the total number of abnormal white matter voxels in each image. However, this approach may not be optimized for accuracy as it counts segmented white matter voxels rather than using a clustering approach, meaning there is less verification during image segmentation of the correct number of white matter regions.

[0010] Another aspect of the present embodiments includes the realization that another notable approach is using an automated proximity-based fiber clustering method. In this approach, the false positives may be filtered out by examining “fiber bundles”, which are grouped integral curves that pass through manually selected ROIs. The fiber bundles for controls may be compared to those of patients, the controls being separated from the patient files, and the controls being automatically labeled as true negatives. However, this approach is again prone to accuracy errors, as it is possible for a patient file to incorrectly be labeled as a control file. One early study in the development of the diagnosis of lung cancer used a two-level convolution neural network to classify images of patients who potentially had lung cancer, attempting to distinguish between the false positive and true positive results. However, neural networks, such as the one used in this study, when appliedto ROC analysis often result in high, but sub-optimal AUROC values. The maximal value in this study was 0.93, signifying room for improvement of the methodology.

[0011] Another aspect of the present embodiments includes the realization that another approach may include the use of a Bloom filter, a special type of artificial neural network that has two layers and takes an autoscaling approach to filtering false positives. Typically, Bloom filters are characterized in ROC analysis by having a perfect true positive rate, and a nonzero false positive rate that depends on the number of elements in the filter. The highlighted study proposes the use of an autoscaling Bloom filter to automatically adjust threshold parameters to achieve an optimal false positive rate. However, the reduction of the false positive rate also causes a slight reduction in the true positive rate, signaling room for improvement.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The various embodiments of the present the present SPM optimizations now will be discussed in detail with an emphasis on highlighting the advantageous features. These embodiments depict the novel and non-obvious SPM optimizations shown in the accompanying drawings, which are for illustrative purposes only. These drawings include the following figures:

[0013] FIG. l is a block diagram illustrating a system for SPM optimization in accordance with an embodiment of the invention;

[0014] FIG. 2 is a block diagram illustrating a client device for SPM optimization in accordance with an embodiment of the invention;

[0015] FIG. 3 is a flow chart illustrating a process for diagnosis of one or more brain conditions using SPM optimization in accordance with an embodiment of the invention;

[0016] FIG. 4 is a flow chart illustrating a process for SPM optimization in accordance with an embodiment of the invention;

[0017] FIG. 5 is a graph illustrating a receiver operating characteristic (ROC) curve of SPM diffusion tensor imaging (DTI) binary classification method for / 2-values (0.001 - 0.05), extent threshold (k = 60), with no activated brain cluster filter in accordance with an embodiment of the invention;

[0018] FIG. 6 is heat map showing single point AUROC value for various extent thresholds and / (-value for cluster size and count without the activated brain cluster filter in accordance with an embodiment of the invention;

[0019] FIG. 7 is a graph illustrating an ROC curve of SPM DTI binary classification method for / ?-values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50% in accordance with an embodiment of the invention;

[0020] FIG. 8 is a heat map showing single point AUCROC value for various extent thresholds and / (-value with the activated brain cluster filter in accordance with an embodiment of the invention;

[0021] FIG. 9 is a graph illustrating an ROC curve of SPM DTI binary classification method for / (-values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50%, for the second cohort of patients in accordance with an embodiment of the invention; and

[0022] FIG. 10 is a heat map showing SPM DTI binary classification method for a second cohort of patients, with / (-values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50% in accordance with an embodiment of the invention.DETAILED DESCRIPTION OF THE DRAWINGS

[0023] The following detailed description describes the present embodiments with reference to the drawings. In the drawings, reference numbers label elements of the present embodiments. These reference numbers are reproduced below in connection with the discussion of the corresponding drawing features.

[0024] The embodiments of the present systems, devices, and methods for SPM optimization are described below with reference to the figures. These figures, and their written descriptions, may indicate that certain components of the apparatus are formed integrally, and certain other components are formed as separate pieces. Those of ordinary skill in the art will appreciate that components shown and described herein as being formed integrally may in alternative embodiments be formed as separate pieces. Those of ordinary skill in the art will further appreciate that components shown and described herein as being formed as separate pieces may in alternative embodiments be formed integrally. Further, as used herein the term integral describes a single unitary piece.

[0025] Turning now to the drawings, systems, devices, and methods for SPM optimization (may be referred to as “SPM optimizations”) in accordance with embodiments of the invention are described. In many embodiments, the present SPM optimizations may overcome the challenges of the Support Vector Machine by not requiring the user to have to manually locate regions of interest (ROIs) on activated brain clusters, as this may be conducted by the Felzenszwalb image segmentation, as further described below. In various embodiments, the invention may therefore be more efficient than Support Vector Machine processes because of the reduced amount of time required defining the ROIs. The process of locating ROIs may be done intrinsically by the Felzenszwalb image segmentation, while in the Support Vector Machine process it must be done manually by a user. The invention may also be superior to other common ROC statistical analysis methods such as, but not limited to, EZ-MAP or “one vs. many” t-tests because the present embodiments employ a clustering approach rather than directly examining white matter voxels. This ensures greater accuracy because there is more verification during the image processing that the activated clusters are in the positions as identified by the algorithms. The invention also overcome the accuracy errors of the fiber bundles based approach, as the control and patient files may already be separated before the processing begins. The invention may also have higher AUROC values compared to many systems that employ a neural network (e.g., some embodiments may reach a maximal value of approximately 0.97 or greater), showing more desirable results to be used for analysis compared to using a fiber bundle based method. Finally, the invention may have no effect on the true positive rate as a Bloom filter does. In some embodiments, the invention may reduce the false positive rate of the control files, and it may do so without affecting the true positive rate.

[0026] In several embodiments, the invention may include the combination of two methods of image processing to achieve a novel functionality of processing patient files such as, but not limited to, MRI patient t-contrast files. For example, the present embodiments may be performed in software (e.g., Python software) as a single code script, combining a function titled “plot_glass_brains”, from Python’s “nileam” module, with a function titled “felzenszwalb”, from Python’s “scikit-image” module. Overall, the inputs to the code may be t-contrast images (e.g., MRI contrast images), and the outputs of the code may be the number of activated clusters detected in a segmented image (the number of areas in the brain that show signs of a disease). The presentembodiments may include the ability to read and process a t-contrast file and count the number and size of the clusters by performing “Felzenszwalb” image segmentation (identifying features within an image or set of images) on the activated brain clusters images. The present embodiments may be utilized for filtering out false negatives in control files in the process of performing SPM for MRI scans. The present embodiments of the invention may be beneficial for AUROC analysis of data from clustered images because the results may be easier to analyze due to the reduced amount of false positives. Systems for SPM optimizations in accordance with embodiments of the invention are further described below.SPM OPTIMIZATION SYSTEMS

[0027] A block diagram illustrating a system for SPM optimizations in accordance with an embodiment of the invention is shown FIG. 1. An SPM optimization system 100 may include subject(s) 102 who may be scanned using imaging device(s) 104, as further described below. In many embodiments, the system 100 may also include a client device 108 and a user 106. In many embodiments, the client device 108 may be utilized to perform processes for SPM optimizations in analyzing scan data captured by the imaging device(s) 104, as further described below. In various embodiments, the system 100 may include any number of imaging devices and / or client devices. In several embodiments, the client device(s) 108 may include various electronic devices, such as, but not limited to, a desktop computer, laptop computer, tablet computer, smart phone, etc. In some embodiments, the SPM optimizations (e.g., SPM processes) may be performed on the client device 108. In some embodiments, the SPM optimizations may be performed on a server 120. In some embodiments, the SPM optimizations may be performed by the client device(s) 108 in conjunction with the server 120. In some embodiments, the SPM optimizations may be performed by the client device(s) 108, the server 120, and / or any other device including, but not limited to, cloud-based devices.

[0028] In reference to FIG. 1, the client device(s) 108 and the imaging device 104 may be connected to, and have access to, the Internet 114 in a manner known to one of ordinary skill in the art. For example, the imaging device(s) 104 may access the Internet 114 using a variety of methods such as, but not limited to, a modem and / or router 110 (and / or using a wireless access point). Further, the client device(s) 108 may access the Internet 114 via a wireless access point112, such as, but not limited to, Wi-Fi (and / or using a modem and / or router). In some embodiments, the imaging device(s) 104 and / or the client device(s) 108 may access the Internet 114 using a cellular network. The system 100 may also include a server 120 in communication with the imaging device(s) 104 and / or the client device(s) 108. In many embodiments, the imaging device 104 may be configured to transmit scan data to the client device(s) 108 and / or to the server 120. In some embodiments, the client device(s) 108 may be configured to run various applications and / or software such as, but not limited to, Jupyter Notebook (e.g., Jupyter Notebook for Python), Matlab, etc. in performing SPM optimizations. In some embodiments, the client device(s) 108 may also include various software and / or environments capable of reading and executing various programing languages, including, but not limited to, Python.

[0029] A block diagram illustrating a client device for SPM optimization in accordance with an embodiment of the invention is shown in FIG. 2. In many embodiments, a client device 108 may be configured for SPM optimizations, as further described below. In some embodiments, the client device 108 may include a processing module 204 that is operatively connected to a display 202 and a communication module 203. The processing module 204 may comprise a processor 206, volatile memory 208, and non-volatile memory 210. In some embodiments, the non-volatile memory 210 may include a client application 211 and / or a SPM application 212 that configures the processor 206 to perform various functions for SPM optimizations, as further described below. For example, in some embodiments, the SPM application 212 may include various applications such as, but not limited to, Jupyter Notebook (e.g., Jupyter Notebook for Python), Matlab, etc.

[0030] In reference to FIG. 2, the client device 108 may receive scan data 214 and / or control data 226 using the communication module 203 from the imaging device(s) 104 and / or the server(s) 120. For example, the scan data 214 may include a patients’ files and the control data 226 may include the control files. In some embodiments, the client device 108 may receive and / or store module packages 216 such as, but not limited to, Nileam, numpy, matplotlib, numpy, pillow, cv2, and scipy, scikit-image Python modules, as further described below. In some embodiments, the client device 108 may also receive and / or store user inputs 215 such as, but not limited to, p and K values 217, values which may determine which sets of t-contrast files 218 will be selected and are used to test different sets of data as part of AUROC data analysis, as further described below. In addition, the client device 108 may also generate, receive, and / or store t-contrast files 218 suchas, but not limited to t-contrast files 218 based on the p and K values 217, as further described below. In some embodiments, the scan data 214 may include the t-contrast files 218, and vice versa.

[0031] In further reference to FIG. 2, client device 108 may generate and store Glass Brain images 220 and gray scale images 222, as further described below. In many embodiments, the client device 108 may conduct segmentation (e g., Felzenszwalb segmentation) and generate and / or store cluster data 224 such as, but not limited to, number(s) and occurrence(s) of unique labels, as further described below. A person of ordinary skill in the art would understand that the data stored in non-volatile memory 210 may be stored in volatile memory 208 as well. In some embodiments, the client device 108 may be in network communication with one or more server(s) (e.g., server 120) in performing any of the processes for SPM optimizations, as described herein. In some embodiments, the t-contrast files 218, Glass Brain images 220, gray scale images 222, cluster data 224 may include patient and / or control data sets. For example, when one or more patients are being diagnosed forbrain injuries, controls may be utilized. For example, the t-contrast files 218 may also include control t-contrast files, the Glass Brain images 220 may also include control Glass Brain images, gray scale images 222 may also include control gray scale images, and the cluster data 224 may also include control cluster data, etc.

[0032] In the illustrated embodiment of FIG. 2, the various components including (but not limited to) the processing modules 204 and the communication module 203 are represented by separate boxes. The graphical representations depicted in FIG. 2 are, however, merely examples, and are not intended to indicate that any of the various components of the client device 108 are necessarily physically separate from one another, although in some embodiments they might be. In other embodiments, however, the structure and / or functionality of any or all of the components of the client device 108 may be combined. In addition, in some embodiments the communication module 203 may include its own processor, volatile memory, and / or non-volatile memory. In addition, the communication module 203 may comprise (but is not limited to) one or more transceivers and / or wireless antennas (not shown) configured to transmit and receive wireless signals such as (but not limited to) satellite, radio frequency (RF), Bluetooth or WIFI. In other embodiments, the communication module 203 may comprise (but are not limited to) one or more transceivers configured to transmit and receive wired signals.

[0033] Although specific systems and devices for SPM optimizations are discussed above with respect to FIGS. 1-2, any of a variety of systems and devices including a variety of imaging devices, client devices, servers, communicating using various communication protocols as appropriate to the requirements of a specific application can be utilized in accordance with embodiments of the invention. Processes for SPM optimization in accordance with embodiments of the invention are discussed further below.SPM OPTIMIZATION PROCESSES

[0034] In many embodiments, a feature of the present invention may include combining two image processing techniques together for creating a filter to remove false positive results from control files of MRI DTI processing. In various embodiments, the present invention may be implemented in software such as, but not limited to Python coding software. For example, a Jupyter Notebook version of Python may be utilized as a platform to write the code.

[0035] A flow chart illustrating a process for diagnosis of one or more brain conditions using SPM optimization in accordance with an embodiment of the invention is shown in FIG. 3. The process 300 may include receiving (302) scan data and / or control data, as further described above. In some embodiments, scan data may include t-contrast files that may include t-contrast images. For example, in some embodiments, the t-contrast files may be MRI patient t-contrast files and the t-contrast images may be MRI contrast images. In some embodiments, the scan data may be processed via image processing (e.g., SPM data processing) to generate the t-contrast files. In some embodiments, the process 300 may also include setting (304) p and K values. For example, in embodiments, the process may receive user inputs such as, but not limited to p and K values, and set (304) the p and K values based on the user input. Further, the process 300 may also include selecting (306) at least one set of t-contrast file. In some embodiments, the p and K values may be used to determine which sets of t-contrast files will be selected (306). Moreover, in some embodiments, the process 300 may also include performing (308) an AUROC data analysis. In some embodiments, the p and K values and / or the selected t-contrast file(s) may be used to test different sets of data as part of an AUROC data analysis, as described further herein.

[0036] In reference to FIG. 3, the process 300 may also include generating (310) one or more Glass Brain images, as further described herein. Generating (310) the one or more Glass Brainimages may include examining the t-contrast files and outputting a maximum intensity image of all of the detected activated clusters in the file. In some embodiments, these images may be converted to gray scale images for better processing. In addition, the process 300 may include performing (312) segmentation such as, but not limited to Felzenszwalb image segmentation. In some embodiments, Glass Brain image(s) and / or a grayscale version of the Glass Brain images may be utilized in the Felzenszwalb image segmentation. In some embodiments, performing segmentation (312) may include take a graphical grid approach to examining an image, labeling each of the clusters based on their size and location. Furthermore, the process 300 may also include generating (314) cluster data. In some embodiments, Felzenszwalb segmentation may assign a unique label to each cluster. To determine the number and size of the clusters in each of the images, the number of unique labels in each image may be counted, as well as the number of times each unique label appears in each image. The number of unique labels in the image is may be an outputted as the total number of clusters, and the number of times each unique label appears may be utilized as the size of each of the clusters. In many embodiments, these numbers may be utilized to filter out false positive results in diagnosis (316) of at least one brain condition (e.g., TBI). In some embodiments, one or more steps in process 300 may be performed on control data (e.g., control contrast files) in diagnosis (316) of the at least one brain condition.

[0037] The various processes described herein may be implemented utilizing code. For example, Python code may be utilized for implementing SPM optimization in accordance with an embodiment of the invention. For example, code implementation of the various processes may include loading in Python modules nilearn, scikit-image, numpy, scipy, pillow, cv2, and matplot.lib that come prepackaged with functionality pertaining to image and mathematical analysis. Functions from these modules may be used throughout the code to conduct the image processing. Following this, the t-contrast files, generated previously using SPM data processing, may be loaded into the Python workspace. These t-contrast files may be fed as inputs into a function (e.g., titled 1). In several embodiments, this “plotting_glass_brains” function may examine the t-contrast files and output the maximum intensity image of all of the detected activated clusters in the file. In some embodiments, these images may then be converted to gray scale images for better processing during the next step. The grayscale images may then be fed into a function titled “felzenszwalb”, part of the scikit-image module, which conducts Felzenszwalb imagesegmentation on the images. Felzenszwalb image segmentation may take a graphical grid approach to examining an image, labeling each of the clusters based on their size and location.

[0038] As further described herein, Felzenszwalb segmentation may assign a unique label to each cluster, so to determine the number and size of the clusters in each of the images, the number of unique labels in each image may be counted, as well as the number of times each unique label appears in each image. The number of unique labels in the image may be outputted from the code as the total number of clusters, and the number of times each unique label appears may be outputted as the size of each of the clusters. These numbers may then be utilized later in the project pipeline to filter out false positive results. In some embodiments, the invention may be part of an effort to find an efficient way to obtain the number and size of clusters in an activated Glass Brain image.

[0039] As further described herein, the processes embodied in software may be used as part of a process in reducing the number of false positives from a set of controls during SPM processing of scans (e.g., MRI scans). In some embodiments, the present embodiments may be used as a supplement to the data processing techniques instead of as a replacement of the overall SPM process for analyzing DTI scans. As further described herein, the software may use two key functions, each from a different published Python module, and combine them for a novel functionality. For example, the Glass Brains function may be from the Python module “nilearn” which can be found on Github at_https: / / github.com / nilearn / nilearn and documentation for the code can be found at_https: / / nilearn. github. io / stable / index.html. Further, the Felzenszwalb function may be from the Python module “scikit-image”, which can be found on Github at https: / / github.com / scikit-image / scikit-image and the documentation for the code can be found at https: / / scikit-image.org / . In addition, Jupyter Notebook for Python can be found at https: / / jupyter.org / .

[0040] In many embodiments, the processes embodied in the software may address the issue of too many false positive results appearing for control files during an AUROC analysis project to optimize SPM processing for DTI files. For example, in certain studies, it may be expected that control files have no false positive results, as a false positive result indicates that the control file is diseased, which should never be the case. Thus, the processes may be used as a method of filtering out false positive results by comparing the number of activated brain clusters in the controls to those in the patients.

[0041] The present invention may also be embodied in and / or utilize various other environments (e g., applications) such as, but not limited to, Matlab. For example, the present embodiments may use Matlab’s image masking application to process created Glass Brain files from Python, then count the number of N4 or N8 connected pixels to determine the number of activated clusters. In some embodiments, this process may require manual input from the user to operate the Glass Brains function and complete the image masking process. This method may also simply count the number of pixels that are connected within an image. Thus, the provided software process may be more time efficient and results in better outputs for what is required for image processing projects. Another software often used for image segmentation tasks may be RIL-contour, which also has implementation into Python Jupyter Notebook. The RIL-contour environment may have more versatility when it comes to annotating medical images, however it typically requires the user to manually define ROIs to be examined. The advantage of using the provided software may be the absence of user input to identify the activated clusters, as the Felzenszwalb Image segmentation function may perform this automatically. Another software method of performing similar analysis on medical images such as the Glass Brain images is ImageI2, a recently developed (in 2017) image analysis program that enables in depth processing of medical images. The provided software may be superior to ImageI2 because it outputs the size in addition to the number of the activated brain clusters. Overall, the data outputted from the provided software process allow for more robust AUROC analysis.

[0042] In many embodiments, the number and size of the activated brain clusters may be outputted from the Python code as .mat variables, where they may be used as part of a filter in Matlab software. In various embodiments, this filter may input these variables and filters out false positive results from the control files based on a percentage threshold.

[0043] A flow chart illustrating a process for SPM optimization in accordance with an embodiment of the invention is shown in FIG. 4. The process 400 may include importing (402) module packages, such as, but not limited to Python module packages. In some embodiments, step 402 may include importing Nilearn, numpy, matplotlib, numpy, pillow, cv2, and / or scipy, sckikit- image Python modules. The process 400 may also include setting (404) p and K values. For example, a user may define p and K values, values which determine which sets of t-contrast files will be selected and are used to test different sets of data as part of AUROC data analysis. Further,the process 400 may include loading (406) t-contrast files. In some embodiments, this may include loading previously generated t-contrast files based on p and K values. The process 400 may also include generating (408) Glass Brain images which may include printing a maximum intensity projection of absolute values of the loaded t-contrast files. Furthermore, the process 400 may include converting (410) the Glass Brain images to gray scale. In some embodiments, converting images to gray scale may allow for greater compatibility and easier functionality with Felzenszwalb segmentation.

[0044] In reference to FIG. 4, the process 400 may include conducting (412) Felzenszwalb segmentation. For example, the gray scale Glass Brain images may be inputted into the Felzenszwalb image segmentation process. Further, the process 400 may include counting (414) number and occurrence of unique labels. For example, the number of different segmentation labels may be outputted as the number of activated clusters, and the occurrence of each unique label may be outputted as the size of each cluster. Moreover, the process 400 may include saving (416) cluster count and cluster size variables. In some embodiments, the number of clusters in an image and the size of the clusters may be saved as separate variables to be used in a filter for AUROC data analysis.

[0045] Although specific processes for SPM optimization are discussed above with respect to FIGS. 3-4, any of a variety of processes and variations of the described processes as appropriate to the requirements of a specific application can be utilized in accordance with embodiments of the invention. An illustrative example in accordance with embodiments of the invention are further described below.ILLUSTRATIVE EXAMPLE - DIFFUSION TENSOR IMAGING AND [18F]-FDG PET VALIDATION

[0046] With an increased need for standardized methodology in accurate diagnosis of traumatic brain injury (TBI), diffusion tensor imaging (DTI) has provided promising diagnostic results as an adjunct modality, but yet remains underutilized. A purpose of this study was to validate the use of DTI with Statistical Parametric Mapping (SPM) for Traumatic Brain Injury (TBI) supporting its use as a diagnostic tool.

[0047] This illustrative example was retrospective and compared controls to patients clinically diagnosed with TBI. Forty-one controls (mean age = 34.1; range, 19 - 58; 28 Males and 13 Females) were screened (n=41) for cognitive impairment and neurological abnormality. Two cohorts, each of eighteen patients (first cohort: mean age, 41.8; range, 23 - 70; 9 Males and 9 Females; second cohort: mean age, 45.7; range, 23-68; 9 Males and 9 Females) clinically diagnosed with TBI (n=36) were pooled. DTI image acquisition was obtained using a 3 Tesla MRI scanner.

[0048] DTI images were analyzed through voxel-based t-tests using SPM, comparing each individual to the normative control group to generate z-maps for each individual control and each individual patient with a TBI. Test statistics were ranged for 2-values (0.001-0.05) and cluster extent values (0, 30, 60, 65, 70, 75). Area underneath a receiver operating characteristic curve (AUCROC) was used to validate diagnostic capability. AUCROC analysis was conducted on all sets of / rvalue and extent threshold values. Significance of results was determined by examining the AUCROC values. A maximal AUCROC of 1 .000 was obtained across the / ?-value range and cluster extent thresholding values specified across the two cohorts. The high AUCROC supports validation of the methodology presented and the use of DTI and SPM as an adjunct diagnostic tool for TBI.Introduction

[0049] TBI remains problematic in diagnosis even with current advances in neuroimaging, as mild injuries may often lack identifiable, structural markers for detection after incident. Common, anatomical imaging techniques used for medical detection such as magnetic resonance imaging (MRI) and computed tomography (CT) carry innate limitations as damage may not only result in focal and diffuse injuries, but also in metabolic dysfunction. Although pragmatic in the context of immediate medical guidance, oftentimes potential sequelae arise from mild trauma that does not show up with primary or secondary injury imaged by anatomical modalities. Thus, even in occurrences where structural injury such as hematoma, hemorrhaging, or edema are not present, TBI may be diagnosed and better managed with functional neuroimaging modalities such as functional MRI (fMRI), positron emission tomography (PET), and advanced and newer structural modalities such as diffusion tensor imaging (DTI). Although not regularly utilized in clinicalsettings, these additional tools provide adjunct methods for determination of TBI and its severity. However, as their popularity and implementation increase, their legitimacy as diagnostic methods remain less established.

[0050] The modalities discussed in this illustrated example include DTI. DTI is an advanced MRI analysis technique that focuses on the translational movement of water molecules. The analysis examines diffusion (Brownian Motion), the tendency of molecules to move at random due to their thermal energy. DTI specifically examines the motion of water molecules along the axons, and is measured along applied gradient X Y Z axes. The diffusion coefficient, D, is written in a 3x3 matrix form (the diffusion tensor) based on the diffusion of water as defined by the 3 axes. Since the random diffusion of water molecules may be constrained by the myelin sheath of the axon, the gradients perpendicular to the axon are constricted relative to the gradient that is parallel with the axon in the normal brain. However, in TBI, damage to the axon may result in reduced constraints of random diffusion of water by the tom myelin sheath and there is less constriction of the gradient perpendicular to the axon. Voxel-based z-map analysis is commonly utilized for tensor estimation followed by Region of Interest (ROI) analysis, manual definition of a priori specific regions of the brain that are of interest for diagnosis of diseases such as TBI.

[0051] As the use of DTI imaging for analysis of TBI increases, there exists a need to verify the discriminant capacity of this modality in conjunction with statistical software as a quantitative diagnostic method. The illustrative example addresses the ability of DTI used with statistical parametric software (SPM; Wellcome Department of Imaging Neuroscience, London, UK; http: / / www.fil.ion.ucl.ac.Uk / e) to recognize significant deviations in water diffusion due to traumatic injury. Analysis is achieved through voxel-based z-map analysis, commonly implemented in clinical research settings utilizing various imaging tools such as MRI, PET, DTI, and various modalities within each technique. Widely used, these voxel-wise tests may be performed to compare two groups of subjects to establish differences between them, such as healthy controls and those suffering from neurocognitive injury or disease. In particular, deviations in water diffusion are mapped through thresholded statistical tests performed within SPM. Areas that show a significant increase or decrease in activity or microstructural parameters (e.g., fractional anisotropy) within a voxel or cluster of voxels are known as activations.

[0052] Also discussed is image segmentation, here utilized to aid the process of analyzing the optimal SPM results with AUCROC analysis. Image segmentation has been employed in various medical analysis studies to assist in the analysis of outputs from imaging modalities, including DTI with TBI. The present embodiments include image segmentation employed as an adjacent tool to aid the SPM analysis.

[0053] Activation detection within SPM-generated contrasts may require setting decision criterion, such as multiple settings for normalization, but more importantly, thresholding parameters to filter the test statistics for significant activations. Two parameters of importance are extent threshold and / 2-value, restricting the cluster size and intensity thresholds. Commonly used in cluster-based extent thresholding, these are set by the user to limit the errors in the test statistics. Once thresholded to limit error rates, the remaining activations may be considered statistically significant, and may be analyzed further. The main aim of this illustrative example is to assess the validity of DTI, when used with SPM, as a diagnostic method of TBI through analysis of a receiver operating curve (ROC).Materials and MethodsSubjects

[0054] Forty-one controls (mean age, 34.1; range, 19 - 58; 28 Males and 14 Females) were screened for cognitive impairment and neurological abnormality. Two cohorts of eighteen patients each (first cohort: mean age, 41.8; range, 23 - 70; 9 Males and 9 Females; second cohort: mean age, 45.7; range, 23 - 68; 9 Males and 9 Females) clinically diagnosed with TBI were pooled.

[0055] The criteria for mild TBI were closed head injury with post traumatic amnesia of less than 24 hour or loss of consciousness for approximately 30 minutes or less and Glasgow coma scale score of between 13 to 15. Moderate TBI had post traumatic amnesia of more than 1 day and less than 7 days or loss of consciousness of 30 minutes to 24 hours and Glasgow coma scale score of 3 to 8. The first cohort was composed of 18 mTBI. The second cohort was composed of 15 mTBI 2 moderate TBI and 1 severe TBI. Image acquisition for patients and controls were acquired using a Siemens 3 Tesla MRI scanner (FWHM, 2.5 mm).SPM Analysis

[0056] DTI images of all subjects were single-subject analyzed through voxel-based morphology using SPM running on MATLAB (Version 7.1.0.246). Two-sample hypothesis testing was performed with each subject against the forty-one controls within SPM8. Age and gender covariates were accounted for within the design matrix, leaving 17 degrees of freedom from 41 images. Linear model parameters were estimated using SPM’s “Classical” option, using the restricted maximum likelihood estimation procedure. Generation of positive and negative t- contrasts were performed with masking, uncorrected / ?-values ranging between 0.001 and 0.05, and extent thresholds of zero (0), thirty (30), sixty (60), sixty-five (65), seventy (70), and seventy- five (75) voxels. The resulting contrasts were analyzed for ROC curve generation.Leave One Out Analysis

[0057] Leave One Out (LOO) Analysis was conducted during initial processing of the control files to ensure no bias in the statistical and data analysis. This process was performed in Matlab (Version 9.9.0.1570001) and was utilized to ensure that the analysis did not compare a control file against a control list containing itself. During generation of two group t-test models, one control scan was always removed from the estimation process. This was done incrementally, so that each file generated by a previous scan was removed from the list used to create the two-group t-test models.Receiver Operating Characteristic (ROC) Generation

[0058] The ROC curve was generated through activation detection, or cluster observation, within the t-contrasts. Thus, the null hypothesis assumes that there are no activations in either control or patient contrasts. A true positive (TP) may be determined by the presence of one or more activations for a subject diagnosed with TBI, while a false positive (FP) may be determined by the absence of clusters for the subject with TBI. A true negative (TN) occurs when the control group does not show any activations, while a false negative (FN) contains one or more activations in this group. Note that determination of a TP or FP may not require both contrasts, positive or negative, to show activation, but may be contingent upon only one displaying an activation. False positive rates (FPR) and true positive rates (TPR) may be calculated from sensitivity and specificity values after t-contrast generation, and plotted to generate the ROC curve.Activated Brain Cluster Filters

[0059] A first test was conducted with the above described procedures. Initial results from this test showed that the inclusion of LOO analysis within the data generation pipeline caused an abnormally high number of control files to register as FP even with strict / -value and extent threshold parameters. To reduce the number of FP in the controls, a third parameter was created and implemented during comparison of t-contrasts. This parameter may be a filter (may also be referred to as an “activated brain cluster filter”) based on the number of activated brain clusters detected in the DTI images in the control and patient t-contrast files. In many embodiments, the activated brain cluster filter may include the present SPM optimizations as described above.

[0060] In various embodiments, the number of the activated brain clusters was calculated in Jupyter notebook (version 4.6) utilizing Python software. First, the maximum intensity projections of the absolute values of the clusters in each t-contrast file were plotted. Felzenszwalb image segmentation was performed on grayscale images of the maximum intensity projections to calculate the number of clusters. The number of clusters was set as the number of different segmentation labels found by the image segmentation process.

[0061] After initial SPM analysis of the control files was conducted, the activated brain cluster filter determined if the following statement was true for each control file: (Number of clusters in current control file) < (Average number of clusters in all patient files * activated brain cluster filter percentage), where the activated brain cluster filter percentage was tested in increments of 10 from 10% to 100% (10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100%). If this statement was true, and the current control file would have been classified as a false positive, the current control file was instead classified as a true negative. For both patient cohorts, an experiment with and without the activated brain cluster filter was conducted.ResultsFirst Data Set Results

[0062] For the first cohort of patients, without the activated brain cluster filter, for / ^-values ranging between 0.001 and 0.05, and an extent threshold of 60, the generated ROC curve is shown (see FIG. 5). A graph 500 illustrating a receiver operating characteristic (ROC) curve 502 of SPMDTI binary classification method for r-values (0.001 - 0.05), extent threshold (k = 60), with no activated brain cluster filter in accordance with an embodiment of the invention is shown in FIG. 5. The graph 500 provides False Positive Rate 504 on the x-axis and True Positive Rate 506 on the y-axis.

[0063] Note that an extent threshold of 60 may not be the optimal extent threshold for this case, but is presented for comparison to the test results from the activated brain cluster filter with an extent threshold of 60. With no activated brain cluster filter, for extent thresholds of 0, 30, 60, 65, 70, and 75, the heat map is shown (see FIG. 6). A heat map 600 showing single point AUROC value for various extent thresholds and / ?- value for cluster size and count without the activated brain cluster filter in accordance with an embodiment of the invention is shown in FIG. 6. The heat map 600 provides extent threshold (K) 602 on the x-axis and p value 604 on the y-axis.

[0064] The single-point area may be calculated based on a single point representative of a specific thresholding combination of extent and - value. The maximal AUROC calculated for an extent threshold of 60 is 0.901 at a y?-value of 0.001. The maximal AUROC across all extent thresholds is 0.913 with an extent threshold of 70 and / r- value of 0.001.

[0065] With the activated brain cluster filter, for ?-values ranging between 0.001 and 0.05, an extent threshold of 60, and an activated brain cluster filter percentage of 50%, the sensitivity and specificity values are shown in the table below. k = 60 Patients (n = 18) Controls (n = 41) / j-value TP FN TN FP Sensitivity Specificity0.001 17 1 41 0 0.94 1.000.002 17 1 41 0 0.94 1.000.003 17 1 41 0 0.94 1.000.004 17 1 41 0 0.94 1.000.005 17 1 41 0 0.94 1.000.006 17 1 41 0 0.94 1.000.007 18 0 41 0 1.00 1.000.008 18 0 41 0 1.00 1.000.009 18 0 40 1 1.00 0.980.010 18 0 40 1 1.00 0.980.050 18 0 11 30 1.00 0.27Table 1. True and predicted classification with calculated sensitivity and specificity for k = 60 with an activated brain cluster filter percentage of 50%.

[0066] TP and FN for the patients may be inversely related as / 2-value changes. As / i-value increases from 0.001 to 0.05, the number of TP increases while FN decreases. While this may suggest a high -value would provide better differentiation, TN decreases and FP increases for the control set as -value increases. Analogously, sensitivity and specificity share an inverse relationship with varying / ?-value. Plotting FPR and TPR, derived from specificity and sensitivity respectively, the ROC curve may be generated for an extent threshold of 60. The maximal AUROC in this case is calculated to be 1.000.

[0067] A graph 700 illustrating an ROC curve 702 of SPM DTI binary classification method for -values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50% in accordance with an embodiment of the invention is shown in FIG. 7. In addition to the cluster extent value 60, additional values of 0, 30, 65, 70, and 75 were analyzed for the same p- value range to determine the effect of extent thresholding on specificity, sensitivity, and diagnostic capability. A single-point area underneath the ROC curve is estimated and shown in the heat map for an activated brain cluster filter of 50%. The graph 700 provides False Positive Rate 704 on the x-axis and True Positive Rate 706 on the y-axis.

[0068] A heat map 800 showing single point AUCROC value for various extent thresholds and / ?- value with the activated brain cluster filter in accordance with an embodiment of the invention is shown in FIG. 8. As illustrated in FIG. 8, the heat map suggests the potential for parametric optimization as certain combinations present a maximal single-point value of one, representative of the ideal diagnostic capability for this methodology in conjunction with the activated brain cluster filter. The heat map 800 provides extent threshold (K) 802 on the x-axis and p value 804 on the y-axis.Second Data Set Results

[0069] For the second cohort of patients, with an activated brain cluster percentage of 50%, the results are shown.

[0070] A graph 900 illustrating an ROC curve 902 of SPM DTI binary classification method for -values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50%, for the second cohort of patients in accordance with an embodiment of the invention is shownin FIG. 9. The graph 900 provides False Positive Rate 904 on the x-axis and True Positive Rate 906 on the y-axis.

[0071] A heat map 1000 showing SPM DTI binary classification method for a second cohort of patients, with -values (0.001 - 0.05), extent threshold (k = 60), and activated brain cluster filter percentage of 50% in accordance with an embodiment of the invention is shown in FIG. 10. The heat map 1000 provides extent threshold (K) 1002 on the x-axis and p value 1004 on the y-axis.Discussion

[0072] The results for the first cohort without the activated brain cluster filter do not fully support use of SPM methodology with DTI, with a maximal AUC value of 0.913. However, focusing on the results that employed the activated brain cluster filter, a maximal AUROC value of 1.000 was achieved at a filter percentage of 50%. The high AUC of 1.000 when the filter was used supports the diagnostic validity of SPM’s usage with diffusion tensor imaging for recognition of TBI. Considering an AUC of 0.5, differentiation using the questioned methodology would be inappropriate as diagnosis would be unable to reject or accept the null hypothesis correctly. As the resulting maximal AUROC is a perfect classifier value of one, the ability to recognize activations due to TBI from contrasts in this case is strongly supported. The high specificities and sensitivities obtained further substantiate the ability of this methodology, as TP and TN can easily be differentiated from false evaluations.

[0073] However, the maximal AUC value of 1.000 in the first cohort, and 0.963 in the second cohort, rely on the relationship between the activated brain cluster filter and its effect on the p- value and extent threshold. The activated brain cluster filter itself is a methodology specific to this illustrative example, and while using LOO with AUROC has been proven a valid method for eliminating study bias, combining LOO analysis with SPM processing remains a relatively unknown practice. However, the usage of the activated brain clusters themselves in the filter as validation of the removal of false positives in the controls suggests that the methodology works well in conjunction with the / ?-value and extent threshold parameters, which are also dependent on the activated brain clusters. Note that the filter itself is meant as an adjunct methodology to help give insight into the optimal / ?- value and extent threshold parameters.

[0074] Diagnostic capacity does not signify high sensitivity and specificity for all parameters tested. Within this trade-off between sensitivity and specificity lies the potential for optimal parametric determination, where ideal parameters may be found to obtain the most differential results for diagnosis. In this set of patients specifically, y-values of 0.007 and 0.008 and an extent threshold of 60 would result in a high sensitivity and specificity of 1.000. Higher / ?-values between 0.009 and 0.050 would instead again result in a sensitivity of 1.000, and specificities between 0.980 and 0.270. In this regard, finding a balance between Type I and Type II error can be achieved through intensity and extent thresholding, as done previously through fMRI studies. Currently, there is no guideline for an objective specification of either value; selection is based on a widely- used / -value < 0.001 as it is arbitrarily more stringent than the more common / ?- value < 0.05. Although there are older standardized protocols for threshold selection, these methods are complex and require additional resources.

[0075] In general, in AUC of .70 or greater has been considered a good classification accuracy with regards to brain imaging analysis.18Thus, a p-value of 0.01 with an extent threshold of 30, resulting in an AUC of 0.939 from the first patient cohort would be considered to have good classification accuracy. Specifically, as tested for this data set and testing procedures, setting an extent threshold value of 30 contiguous voxels suggests that / ^-values of either 0.007 or 0.008 have the highest diagnostic capability for an extent threshold of 60, as the FPR of 0 and TPR of 1 result in an AUC of 1.000. For larger cluster extent thresholds of 65, 70 and 75, a / >-value of 0.007 also resulted in an AUC of 1.000. A / i-value of 0.008 and 0.009 also resulted in an AUC of 1.000 for extent thresholds of 70 and 75 and a / - value of 0.01 resulted in an AUC of 1.000 for an extent threshold of 75.

[0076] lust as there is no objective guideline for / ?-value. cluster extent thresholding remains arbitrarily selected based upon a posteriori experiments. Similar to possible -value optimization, extent threshold evaluated at different cluster sizes provides varying TPR and FPRs for each p- value. Based on a single-point analysis for each ROC curve generated with a constant extent threshold and varying / ?-value, a smaller extent threshold generally may show a larger AUC for stricter / ^-values. The balance in this case lies between localization power within an increasingly larger contiguous set of voxels, and a decreasing significance within a relatively precise activation.Considerations

[0077] The study used a small number (n=18) of patients in both cohorts in comparison to the number of controls (n=41) that were used for comparison during the study. The results, including the maximal AUCROC value of 1.000, depend on the activated brain cluster fdter, a methodology specific to this illustrative example has not been explicitly tested or verified by previous AUCROC analysis studies. Additional verification of the methods used in this illustrative example would further validate the results presented.Conclusions

[0078] Use of DTI with SPM as a diagnostic method is supported based on the optimal AUC value of 1.000 obtained for the selected extent and intensity thresholding values. The determination of possible optimal thresholding parameters specific to methodology for TBI may be seen with the heat map generated for various combinations of extent threshold and -value parameters. Possible case-specific evaluation of extent and intensity thresholding based on clinically validated assessments could be further validated based upon a larger cohort of patients and future evidence. Additional testing of a filter such as, but not limited to, the activated brain cluster filter could also further validate the optimal extent and intensity thresholds as determined by this study. In the same light, variation to methodology would warrant additional validation studies as minute changes may alter classification capability.

[0079] Although a specific illustrative example of SPM optimizations is discussed above with respect to FIGS. 5-10, any of a variety of experiments and evaluations for SPM optimizations as appropriate to the requirements of a specific application can be utilized in accordance with embodiments of the invention. While the above description contains many specific embodiments of the invention, these should not be construed as limitations on the scope of the invention, but rather as an example of one embodiment thereof. It is therefore to be understood that the present invention may be practiced otherwise than specifically described, without departing from the scope and spirit of the present invention. Thus, embodiments of the present invention should be considered in all respects as illustrative and not restrictive.

Claims

WHAT IS CLAIMED IS:

1. A non-transitory machine-readable medium of a client device storing a program for statistical parametric mapping (SPM) optimization, the program executable by a processor of the client device, the program comprising instructions, which when executed cause the client device to: select at least one set of t-contrast fdes from a patient’s contrast files; generate at least one Glass Brain image using the at least one set of t-contrast files, wherein the at least one Glass Brain image includes a three-dimensional (3D) visualization of the patient’s brain; perform a segmentation process on the at least one Glass Brain image, wherein the segmentation process generates at least one segmented image; generate cluster data, wherein the cluster data includes data for at least one activated cluster; and diagnose a brain condition for the patient using the cluster data.

2. The non-transitory machine-readable medium of claim 1, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to receive scan data comprising the patient’s contrast files.

3. The non-transitory machine-readable medium of claim 2, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to set p and K values.

4. The non-transitory machine-readable medium of claim 3, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to set the p and K values by receiving the p and K values from a user.

5. The non-transitory machine-readable medium of claim 3, wherein the p and K values are used to select a set of control contrast files and the at least one set of t-contrast files for an area under the receiver operating characteristic curve (AUROC) data analysis.

6. The non-transitory machine-readable medium of claim 3, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to select the at least one set of t-contrast files based on the p and K values.

7. The non-transitory machine-readable medium of claim 1, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to generate the at least one Glass Brain image by using a maximum intensity projection of absolute values of the at least one set of t-contrast files.

8. The non-transitory machine-readable medium of claim 7, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to convert the at least one Glass Brain image to gray scale.

9. The non-transitory machine-readable medium of claim 1, wherein the segmentation process is a Felzenszwalb segmentation process.

10. The non-transitory machine-readable medium of claim 9, wherein the Felzenszwalb segmentation process labels each cluster of the at least one activated cluster based on a size and a location associated with the at least one activated cluster.

11. The non-transitory machine-readable medium of claim 9, wherein the Felzenszwalb segmentation process assigns a unique label to each cluster of the at least one activated cluster.

12. The non-transitory machine-readable medium of claim 11, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to determine a number of activated clusters by counting a number of unique label in each of the at least one segmented image.

13. The non-transitory machine-readable medium of claim 11, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to determine a size of each cluster of the at least one activated cluster by counting a number of times each unique label appears in each of the at least one segmented image.

14. The non-transitory machine-readable medium of claim 12, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to receive control data comprising control contrast files.

15. The non-transitory machine-readable medium of claim 14, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to: select at least one set of control t-contrast files from the control contrast files; generate at least one control Glass Brain image using the at least one set of controlt-contrast files, wherein the at least one Glass Brain image includes a 3D visualization of a control brain; perform the Felzenswalb segmentation process on the at least one control Glass Brain image, wherein the Felzenswalb segmentation process generates at least one segmented control image; and generate control cluster data, wherein the control cluster data includes data for at least one activated control cluster.

16. The non-transitory machine-readable medium of claim 15, wherein the Felzenszwalb segmentation process assigns a unique control label to each cluster of the at least one activated control cluster.

17. The non-transitory machine-readable medium of claim 16, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to determine a number of activated control clusters by counting a number of unique control label in each of the at least one segmented control image.

18. The non-transitory machine-readable medium of claim 17, wherein the program comprises further instructions that, when executed by the processor, further cause the client device to diagnose the brain condition using the cluster data by filtering out false positives.

19. The non-transitory machine-readable medium of claim 18, wherein filtering out false positives includes comparing the number of activated control cluster to the number of activated cluster.

20. The non-transitory machine-readable medium of claim 18, wherein filtering out the false positives includes removing the false positive results from controls files of a magnetic resonance imaging (MRI) diffusion tensor imaging (DTI) process.

Images