A device for automatic schizophrenia disease detection using FMRI images and method thereof

Abstract

ABSTRACT A DEVICE FOR AUTOMATIC SCHIZOPHRENIA DISEASE DETECTION USING FMRI IMAGES AND METHOD THEREOF The present invention relates to the field of artificial intelligence in healthcare monitoring. Particularly, the present invention relates to the design, development, and deployment of a representation-based learning system for the automatic detection of schizophrenic patients using pre-processing of fMRI images. The representation based learning system of the present invention incorporates an fMRI monitoring system operatively coupled with an artificial intelligence based controller having a FPGA based processor embedded with a convolutional neural network module. Figure 1

Description

BACKGROUND OF THE INVENTION

[002] Schizophrenia (SZ) is a neurological disorder caused by hallucinations,cognitive dysfunctions, and delusions that can severely affect the patient's brain functionsand daily activities. About 21 million people around the world have SZ, which usuallystarts in adolescence or early adulthood. It reduces the life expectancy of patients by 15years. Manual screening of SZ using fMRI scan images is prone to human error, timeconsuming,and challenging due to image contamination from a variety of factors. Motionartifacts in fMRI scans is a challenging problem.

[003] According to the World Health Organization (WHO), the average age of onsetof this disorder is 18 in women and 25 in men, respectively, while the prevalence rate ofthis disorder is higher among men. According to the National Alliance on Mental Illness(NAMI), 1% of Citizens of the United States of America (USA) suffer from SZ. As perthis organization, SZ begins in men's twenties and women are thirteen's. It is difficult todiagnose SZ because of its heterogeneity this mental disorder lacks specific effectivebiomarkers. To diagnose SZ, a few clinical symptoms, including physical, psychiatric,and psychological indicators, need to be evaluated.

[004] As far as scientists are concerned, diagnosing SZ and its related disorders hasalways been challenging. The treatment of SZ is complicated for physicians.Neuroimaging modalities are the method to diagnose SZ and are divided into functionaland structural modalities. These data provide specialists with essential informationregarding the function and structure of the brain for the diagnosis of SZ. Structuralneuroimaging modalities include two methods of Structural Magnetic Resonance Imaging(SMRI) and Diffusion Tensor Imaging (DTI), which show the structure of the humanbrain and its structural connectivity, respectively, owing to their high spatial resolution.Structural neuroimaging modalities based on MRI can visualize white matter, graymatter, and CSF tissues in the brain, as well as explore abnormalities.

[005] Functional neuroimaging modalities for the diagnosis of SZ includeelectroencephalography (EEG), Magneto Encephalography (MEG), and functionalneuroimaging modalities. It has not been used in a large number of studies because of itslow informational resolution. A few of the advantages of MEG for the diagnosis of SZinclude accurate display of brain function during a disorder, accurate source localization,and high resolution.

[006] Reference can be made to US Patent No. US 8,870,737 B2, that describesmethods, devices, and systems for a novel, inexpensive, easy to use therapy for a numberof disorders. Described are methods and devices to treat disorders that involve nomedication. Methods and devices described herein use alternating magnetic fields togently "tune" the brain and affect mood, focus, and cognition of subjects.

[007] Another US Patent No. US 6,978,179 B1 discloses a system for detecting asimplified magnetic brain wave stimulation device gets electrical impulses from pre20recorded brain wave patterns. The user's brain wave patterns trigger electromagnets ontheir neck near their brainstem to generate magnetic pulses that impact their brain centersto achieve a desired mental state. Recorded, recognized, divided, and repeated rhythmicelectrical brain wave patterns are simplified to generate a particular mental state. Recordthe simplified patterns on a suitable memory medium. When the memory means isactivated, rhythmic electrical signals are sent to electromagnets near the user's neck forelectromagnetic stimulation of the brain stem, influencing desired mental states. Amodest magnetic stimulation is felt but not seen by the user. To affect mental state, mildmagnetic stimulation travels via the brain stem to the cerebral centers.

[008] However, such prior EEG systems can only identify EEG problems withfrequency utilization and cannot continuously monitor SZ patients using fMRI images.

[009] Accordingly, there is a need to develop a non-invasive continuous monitoringtechnique to monitor the health and comfort of an individual suffering from SZusingfMRI images.

[0010] Given the aforementioned issues pertaining to the current state of the field, itis essential to devise a system that maintains ongoing surveillance of individualsdiagnosed with schizophrenia (SZ). This system should possess the capability toautonomously identify patients by monitoring alterations in blood flow via the use offMRI images.OBJECTIVES OF THE INVENTION

[0011] The primary objective of the present invention is to provide automaticdiagnosis of schizophrenia disease with FPGA based hardware using an fMRI dataset.

[0012] Another objective of the present invention is to provide pre-processing of thefMRI dataset using statistical parametric mapping (SPM) and realigning of fMRI imageswith a reference image.

[0013] Yet another objective of the present invention is to provide enhance modelsthat are based on convolutional neural networks (CNNs), which have remarkablepotential for providing an accurate and effective system for identifying schizophrenia.

[0014] Yet another objective of the present invention is to provide an FPGA-basedsystem to find the test pattern and the number of times the test image reflects in themodel.

[0015] Yet another objective of the present invention is to enhance a hardwareecosystem with pre-verified reference designs and an enhanced memory interface with asystem-on-chip (SOC) test package for Xilinx Zynq-7000.

[0016] Yet another objective of the present invention is to reduce the processing timeby creating an FPGA-based solution that includes a decreased pre-processed featurevector for the categorization job, as well as an acceleration platform with a customizableapplication methodology.SUMMARY OF THE INVENTION

[0017] The present invention relates to the field of artificial intelligence in healthcaremonitoring. Particularly, the present invention relates to the design, development, anddeployment of a representation-based learning system for the automatic detection ofschizophrenic patients using pre-processing of fMRI images. The representation basedlearning system of the present invention incorporates an fMRI monitoring systemoperatively coupled with an artificial intelligence based controller having a FPGA basedprocessor embedded with a convolutional neural network module.

[0018] The FPGA board of ZC706 schizophrenia identification based on deeplearning offers numerous helpful features for those with schizophrenia. The mobilehardware's deep learning system based on FPGA enables uncomplicated disease detectionfor patients. This board makes the system's specifications easy to modify, making itsignificantly more effective than previous systems. Due to their adaptability, FieldProgrammable gate arrays (FPGAs) have been utilized as hardware accelerators forneural network models. By utilizing the parallel processing capabilities of FPGAs, thequantity of MAC (multiply-accumulate) operations conducted during training wassignificantly reduced. The FPGA-based solution was 15.3 times quicker and required 60times less data dependence reduction than the GPU implementation. Pre-processing offMRI image improves system accuracy and this innovation demonstrates that aclassification algorithm based on a convolutional neural network (CNN) can be trained toaccurately label fMRI scans for schizophrenia. Types of datasets, their modalities, preprocessingapproaches, pre-processing processes, and toolboxes for feature extraction,feature selection, and feature reduction categorization; essential information for AI-basedSZ diagnosis. This innovation enables technical professionals to control their systemsremotely. In addition, only one fMRI image is required to operate this device.BRIEF DESCRIPTION OF DRAWINGS

[0019] The present invention can be comprehended in its entirety by referring to theadjacent illustrations in conjunction with their detailed description, which include:

[0020] Figure 1 is a block diagram illustrating detection of pattern using FPGA basedimage processing of fMRI images.

[0021] Figure 2 illustrates a comprehensive block design of the hardware basedschizophrenia classification system.

[0022] Figures 3A & 3B illustrate simulated results generated using a SPM tool.

[0023] Figure 4 illustrates a hardware based CNN model specifically utilized forclassification purpose.

[0024] Figure 5 demonstrates hardware parallel processing in accordance with someembodiments.

[0025] Figure 6 illustrates hardware components in accordance with someembodiments.DETAILED DESCRIPTION OF INVENTION

[0026] The following detailed description and embodiments set forth herein below aremerely exemplary out of the wide variety and arrangement of instructions which can beemployed with the present invention. The present invention may be embodied in otherspecific forms without departing from the spirit or essential characteristics thereof. All thefeatures disclosed in this specification may be replaced by similar other or alternativefeatures performing similar or same or equivalent purposes. Thus, unless expressly statedotherwise, they all are within the scope of the present invention.

[0027] Accordingly, those of ordinary skill in the art will recognize that various changesand modifications of the embodiments described herein can be made without departingfrom the scope of the invention. In addition, descriptions of well-known functions andconstructions are omitted for clarity and conciseness.

[0028] The terms and words used in the following description and claims are not limitedto the bibliographical meanings but are merely used to enable a clear and consistentunderstanding of the invention. Accordingly, it should be apparent to those skilled in theart that the following description of exemplary embodiments of the present invention areprovided for illustration purpose only and not for the purpose of limiting the invention.

[0029] It is to be understood that the singular forms "a", "an", and "the" include pluralreferents unless the context clearly dictates otherwise.

[0030] The foregoing objects of the invention are accomplished and the problem andshortcoming associated with the prior art, technique and approaches are overcome by thepresent invention as described below the preferred embodiments.

[0031] Figure 1 of the present invention illustrates a block diagram describing thedetection of a pattern using FPGA based image processing of fMRI images. The blockdiagram of the representation based learning system incorporates pre-processed fMRIimages coupled with an artificial intelligence based controller having a FPGA based Armcortex-A9 processor embedded with a convolutional neural network module. The blockdiagram describes the process of detection by providing one or more features andelements in accordance with the disclosed embodiments: preparing images for input at101, processing of images at 102, pixels value of pre-processed image at 103, hardware(FPGA) based CNN's model at 104, convolution of trained images at 105, convolution oftested images at 106, detection of final pattern at 107, and identifying two classificationsby Softmax at 108.The features that are described and / or illustrated with respect to one ormore embodiments may be used in the same way or in a similar way in one or more otherembodiments and / or in combination with or instead of the features of the otherembodiments.

[0032] Figure 2 illustrates a comprehensive block design of hardware basedschizophrenia classification system. In an embodiment of the present invention, anFPGA-based detection system for detecting schizophrenia disease in a patient is provided.The system provides preparing a test image for those who may have schizophrenia andfurther providing the test image to a trained CNNs based model and said model furtherimplements a schizophrenia dataset, for example, a dataset provided by Center forBiomedical Research Excellence (COBRE). It may be appreciated by person skilled inthe art that COBRE dataset is used for comparing normal versus diseased classificationusing a dataset that is a part of International Neuroimaging Data-Sharing Initiative under1000 Functional Connecters Project, comprising fMRI data of the 74 healthy subjects(controls) and the 72 subjects of Schizophrenia with varying ages ranging from 18 to 65years in both the classes.

[0033] In some embodiments, fMRI data is pre-processed using SPM (StatisticalParametric Mapping). The pre-proceed data images are converted into pixels with theMatlab tool. The converted pixels in binary format are then provided to the CNN-basedhardware, which uses block random access memory (BRAM) to store all values. It is thenserially transferred to a FPGA board. The Hardware model is comprised of plurality ofsub-modules, for example, Convolution layer, Filters, Adders, max-pooling layers, andSoftmax function. The sub-modules perform functions in a sequence or pipeline mannerwith a clock and a reset is applied to the test bench.

[0034] In some embodiments, the data flowing through the multiplier and addermodules is related to the pixel values of the training images and the test images. Toconvolute train images and test images, a 3x3 filter with one stride is used. The numberof convolution operations is dependent on image size and filter size.

[0035] In some embodiments, a SPM tool may be utilized for pre-processing of fMRIimages. SPM is statistical method for determining differences in recorded brain activityby building spatially protracted analytical processes that is used to test hypotheses duringreal-world neuro imaging research and the application of neuro psychologicaltechnologies. The SPM sequentially analyses provided brain data and segregates a timeseries of collected images from identical subjects. It exemplifies the application of SPMtheoretical ideas in a practical way through its all-inclusive evaluation tool. TemporalProcessing involves slice timing operations, and Spatial Processing involves realignment,co-registration, segmentation, normalization, and smoothing in time series. The fMRI preprocessinvolves, slice timing, realignment, co-registration, segmentation, normalizationand de-noising. In an embodiment, the SPM tool may be a SPM12 based software.Figures 3A & 3B illustrate simulated results generated using SPM12 software.

[0036] The present invention is a representation-based learning system for SZdetection that includes an fMRI image recording pre-processing of users and an artificialintelligence-based controller that is configured to receive and process fMRI images. In anembodiment, the artificial intelligence-based controller may include plurality of input andoutput ports, one or more processors connected to a memory unit, and the processorincludes an image processing module configured to convert 2-D images to pre-processingimages, as well as a convolutional neural network module configured to distinguishbetween normal and SZ patients.

[0037] In another embodiment, the image processing module and the convolutionalneural network module are operationally linked. In yet another embodiment, aconvolutional layer module is configured to apply a 2D-convolution operation to the inputimage in order to extract deep features from the data; a pooling layer module isconfigured to reduce the spatial dimensions of the output produced by the convolutionlayer module; and a fully connected layer module is configured to convert the 2D outputfeature map into 1D.

[0038] In order to calculate the slice timing, data images are placed with frames sizeof 146. The first volume of the time series has voxel that is a volume element, also knownas a voxel, is a three-dimensional square or rectangular portion of brain tissue thatcorresponds to each pixel dimensions of 64x64x32, wherein 32 is the number of zdimensionalslices. In an embodiment, a Repetition Time (TR) and an acquisition period(TA) are provided, wherein the TR is 2 seconds, and the TA is determined by using theTR and number of slice parameters. The slice order is sequential ascending order that is[1:2:32 2:2:32] and 1 for the reference slice. Realignment calculating the six translationparameters (right, forward, up, in mm) and rotation parameters (pitch, roll, and yaw, inradians) of the volumes collected for the subject with respect to a reference less than 3mm motion are taken into consideration in our model. Grey matter and cerebrospinal fluid(CSF) images are all produced from a single image during the segmentation process.

[0039] In co-registration, the functional data and anatomical data are aligned based onreference images and source images, generally with the functional data being used asreferent images and the anatomical data being used as the source images.

[0040] In some embodiments, the normalization of all data to standardized space andaligned the brains outlines and internal structures. As illustrated in Figure 3B, the subjectis further examined with mean functional and anatomical images, and its ventricles arealigned well. Segment the anatomical images in order to align the difference tissue types.The segmentation illustrates the activation of each element. Normalize all data tostandardized space and aligned the brains outlines and internal structures. General outlineif both anatomical and functional data are aligned with the MNI template, this showedthat the pre-processing was carried out successfully. The BOLD signal sometimes retainsnoise or non-neural variability after preprocessing owing to physiological, outlier, andresidual subject-motion effects.

[0041] Figure 4 illustrates a hardware based CNN model specifically utilized forclassification purpose. Conventional machine learning methods need a significant amountof manual characteristic preprocessing, deconstruction, selection, and classification.Typical machine learning, even after numerous trials, the predicted model performance isnot achieved. A modern branch of machine learning called deep learning searches formultiple interconnected representational layers. Convolutional neural networks areeffective DLMs used for a variety of applications, including automatically classifying,object detection, and segmentation, as they have the highest classification performance. Itis impact by the organic brain system. The CNN architecture has input layer, output layer,numerous hidden layers, and in addition to the softmax layer for classification. Eachlayer's objective is to process the data that was obtained by the preceding layer. To locateand extract the characteristics from the hidden layers, several convolution layers (CLs)and pooling layers (PLs) are utilized. A combination of CL and PL is employed toproduce high-level feature extraction. Fully connected layers (FCLs) and softmax layersare utilized in order to categorization. This sub-section provides an overview about theway every component operates. A CNN module is a collection of well-organizedprocesses that allows a network to be fed with relevant input and learn the appropriaterepresentations for classification automatically. It is programmed to automatically andadaptively learn the spatial hierarchy of features through back-propagation by utilizingseveral blocks, including convolutional layer (CL), pooling layer (PL), and fullyconnected layer (FCL). The CL and PL extract features while FCL map the extractedfeatures.

[0042] Convolutional layer (CL) module: CL is the primary building block of aCNN. It applies a 2D-convolution operation to the input image, which helps to extractdeep features from the data. The convolution is performed with a set of learnable filters,which are also called kernels or weights. These filters scan the input by moving across it(called stride), evaluating "dot" products between the kernel values and input at eachposition.

[0043] Mathematically, a 2D-convolution operation is given as: b) B(a + i, b + j)This process results in a new set of 2D feature maps, which are then passed on to the nextlayer in the network.

[0044] The output feature map dimension (Fwo*Fho *Fco) using a k*k convolutionalkernel, with padding of p, and stride of st is computed as:where Iwi * Ihi* ci represents the size of an input image with ci number of channels.Convolutions help to reduce the size of the data and also allow the network to learnspatial hierarchies of features. Activation functions are used in convolutional layers tointroduce non-linearity into the network. Without non-linear activation functions, a neuralnetwork composed of just linear layers (like convolutional layers) would be equivalent toa single linear layer and would not be able to learn complex, non-linear relationships inthe data. Some examples of activation functions include the Rectified Linear Unit(ReLU), which replaces all negative values with 0, and the sigmoid function, whichsquashes input values into the range of 0 and 1.ReLU activation is widely used in CNNsbecause it helps to alleviate the vanishing gradient problem, which can occur when usingactivation functions like the sigmoid function. The ReLU activation function iscomputationally efficient and helps to improve the performance of the network byintroducing non-linearity and allowing the network to learn more complex features.Consequently, the ReLu activation function is employed as: ReLU(x)=max(0,x).

[0045] Pooling layer (PL) module: The PL is typically used after a Convolutionallayer (CL) and it is employed to reduce the spatial dimensions of the output produced bythe convolution layer CL. This dimensionality reduction is also known as down-sampling.Several types of pooling operations can be used, such as max pooling and averagepooling. In max pooling, the maximum value in the pooling window is selected, and inaverage pooling, the average value in the pooling window is selected. The poolingwindow is typically a square window of a fixed size, and it is moved across the featuremaps in strides. At each position, the values in the window are pooled together accordingto the chosen pooling operation, resulting in a new set of feature maps with smallerdimensions. Pooling operation help to reduce the computational cost and the number ofparameters in the network, while also providing some robustness to small translations anddeformations in the input. It also helps to make the features learned by the network moretranslational invariant.

[0046] Fully connected layer (FCL) module: Fully-connected means that eachneuron in the upper layer is connected to every neuron in the next layer. FCL is a standardartificial neural network, with each input connected to each output by a learning weight. Itworks similarly to a standard multi-layer perceptron neural network. It transforms the 2Doutput feature map into 1D. The number of output classes decides the total number offully-connected neurons in the last layer. To assign the output value from the last FCL tothe desired class probability, a Softmax function is utilized. Each value ranges between 0and 1.

[0047] Softmax function module: The softmax function, also known as thenormalized exponential function, is a mathematical function used in various machinelearning and statistical applications, particularly in the context of multi-classclassification problems. It is used to convert a vector of raw scores (often called logics)into a probability distribution over multiple classes. The softmax function is commonlyused in neural networks, especially for the final layer when classifying multiple classes.

[0048] In an exemplary embodiment, the CNN module incorporates four CLs, threePLs, and two FCLs. The architecture and number of trainable parameters of the saidmodule can be tuned by adding and dropping several layers, depending on theapplication.

[0049] In a preferred embodiment, the method of determining the health patient ofthe SZ patient via the system of the present invention includes:- MRI images are recorded using a Magnetic Resonance Imaging (MRI)machine, which generates detailed images of the brain's structure and function,including traditional closed-bore MRI machines and more open and accessibledesigns.- Brain activity is measured by detecting changes in blood flow.- This approach is based on the idea that cerebral blood flow and neuronalactivity are inextricably linked.- When a part of the brain is used, blood flow to that part of the brain rises.

[0050] Figure 5 demonstrates hardware parallel processing. The diagram of thepipeline structure of the system to speed up classification time in retrospect to thesoftware simulations; a hardware CNN classification pipeline was designed. Acustomized hardware solution was investigated which contained the ZC706 Xilinx FPGAIP core to control the data being fed into the classifier. Overall system design where theclass output presents the result of the classification process. After both images have beenread and classified, 1 is output from the output port. The proposed hardwareimplementation includes an IP core that takes the inputs loaded into the local memoryfrom the block memory generation (BRAM) which helped in reducing the time taken fordata to be loaded to the classifier.

[0051] Figure 6 illustrates hardware components in accordance with the disclosedembodiments of the present invention. In an embodiment, a ZC706 Evaluation Board Kitfor Xilinx Zynq-7000 SoC is provided that offers a hardware platform for designing andtesting applications intended for the Zynq-7000 XC7Z045 SoCs. It may be appreciated bythe person skilled in the art that ZC706 Evaluation Board may include one or morefeatures generally found in various embedded processing systems, for example, twoUART interfaces, general-purpose I / O, an Ethernet PHY, DDR3 SODIMM andcomponent memory, and a four-lane PCI Express interface with dual-arm Cortex-A9 coreprocessors to enable serial communication. Such communication can occur via Input / Output (I / O) interfaces. In another embodiment, a FPGA Mezzanine Card interface isprovided to increase input and output.

[0052] The present invention may be implemented using general-purpose or specificcomputing or processing units, or a combination thereof. A computing or processing unitmay include a computer system or server that may include, but are not limited to, one ormore processors, memory including volatile and non-volatile memories, plurality ofinterface ports, etc. Different wired or wireless communication methods may beimplemented for providing connectivity between disclosed system and other local orremote devices or systems, such as a local area network (LAN) based Ethernet (10-100-1000 Mbps), a wireless wide area network (WAN), or a public network (Internet).

[0053] In a preferred embodiment, the present invention discloses a device forautomatic schizophrenia disease detection in a subject patient using fMRI imagescomprising:plurality of field programmable gate array (FPGA) using convolution neural network(CNN) and a processor configured to execute instructions to perform operations,comprising:performing data analysis on a data comprising plurality of fMRI images of thesubject using a Statistical Parameter Mapping (SPM), wherein plurality of said data ispre-processed with the SPM and a Gaussian filter for realigning one or moreindividual volumes together;estimating the value of each individual volume which is out of alignment withrespect to a reference volume, wherein the reference volume is set in a field whichallows specifying whether the one or more volumes aligned to a mean of all thevolumes;analyzing one or more types of data including a realignment data, a rotation data,and a translation data for measuring change in the movement of the subject, whereinif the subject moved beyond the size of a voxel, or a sudden spike of more than half ofthe voxel size is detected, it is removed for further processing;aligning a functional data to an anatomical data, the functional data comprises oneor more reference images and the anatomical data comprises one or more sourceimages, wherein the reference image remains stationary and the source image isarranged to move around until a best fit is found between the reference image and thesource image;examining the subject with mean functional and anatomical images to identify thealignment in ventricles thereof;segmenting the anatomical images in order to align the difference tissue types toillustrate the activation of each element and normalizing all data to standardized spaceand further aligning brains outlines and internal structures of the subject; andsmoothing the data by combining one or more signals and canceling noise byusing spatial smoothing with Gaussian kernel Intensity technique.

[0054] In another embodiment, the device comprises a realignment data that furtherprovide movement of the subject in a three dimensional coordinate axes space, therotation data indicates rotation of the subject around each coordinate axis, and thetranslation data provides movement of the subject along each coordinate axis.

[0055] In yet another embodiment, the device comprises a CNN-based hardwareautomatically detects and classifies fMRI images.

[0056] In yet another embodiment, the device comprises pre-processing of fMRIimages that utilizes the SPM technique along with Gaussian filter.

[0057] In yet another embodiment, the SPM technique is SPM12.

[0058] In yet another embodiment, the FPGA-based hardware is reconfigurable andfast due to its parallel processing.

[0059] In yet another embodiment, a Xilinx FPGA ZC706 valuation board is used tosynthesize hardware architecture to compare with the CNN based software to findaccuracy.

[0060] In yet another embodiment, the FPGA is connected using a serial interconnectand a universal asynchronous receiver-transmitter port.

[0061] In yet another embodiment, the MRI images are collected by The Centre forBiomedical Research Excellence (COBRE) data centre.

[0062] In yet another embodiment, the subject data is stored in on-chip memory of theFPGA, as well as in a secondary memory for further analysis.

[0063] In another preferred embodiment, the present invention discloses a method forautomatically detecting schizophrenia disease in a subject patient using functionalMagnetic Resonance Imaging (fMRI) images, comprising:performing data analysis on a data comprising plurality of fMRI images of the subjectusing a Statistical Parameter Mapping (SPM) by pre-processing individual subject data,wherein pre-processing includes motion correction, normalization, filtering, slice timingcorrection, artefact removal, and Gaussian filter smoothing of said data;identifying differences between a Normal group and a SZ group by evaluating data fromboth the groups using one or more statistical models, and the data is interpreted byvisualizing and reporting group-level brain activity variations between the Normal and SZgroups;estimating the deviation between each individual volume and a reference volume,wherein a parameter of the reference volume specifies whether the volumes align to themean of all volumes or to the first volume;analyzing realignment of subject movement in x, y, and z dimensions, wherein arotation of the subject is measured around each axis, and a translation of the subjectmovement is measured along each axis;eliminating a subject from further analysis if the subject went beyond a voxel or asudden spike larger than half of the voxel is measured;aligning a functional data to an anatomical data using two fields: a first field includesone or more reference images and a second field includes one or more source images,wherein the functional data is used as a reference image and the anatomical data is usedas a source image such that the reference image remains stable while the source image ismoved to get the greatest match;examining the subject with mean functional and anatomical images;using anatomical picture segmentation to align different tissue types, wherein thesegmentation shows how each element is activated, and standardize the data and furtheralign brain contours and architecture; andsmoothing the data by averaging one or more signals and cancelling noise using aGaussian kernel by increasing the data intensity.

[0064] In another embodiment, the method provides that the Function connection,SBC, ROI, and second-level analysis are conducted before providing to the FPGA basedCNN

[0065] The advantages of the present invention are discussed herein:- The developed system is suitable for diagnosis of schizophrenia. It is the firstsystem to use a ZC706 FPGA evaluation board to diagnose schizophrenia usingfMRI.- The presented framework takes preprocessed fMRI images, and SPMimproves system performance by reducing motion artifacts and increasing imagesignal to noise ratio.- In comparison with the previous systems, they are significantly more useful,due to their flexibility in implementing and modifying neural network models.- FPGA has been used as a hardware accelerator to improve kernel performancethat is 3x3 in size.- MAC operations (multiply-accumulate) were reduced during training byutilizing the parallel processing capabilities of FPGAs.- The FPGA-based solution offered a 15.3-time speed-up over the GPUimplementation with 60 times less data dependency reduction.- The created hardware system is synthesized, implemented, and produced bitstreams, allowing it to be conveniently coupled with FPGA and also supporting theinvention's realization using application specification integrated circuit (ASIC)approach.

[0066] The present invention utilizes the benefits of Convolutional Neural Network(CNN) model for the classification of normal and schizophrenia brain patterns, leveragingthe immense computational capabilities of a Field-Programmable Gate Array (FPGA)based AI solutions in the realm of mental health diagnostics.

[0067] High Classification Accuracy: The CNN model, when implemented on theFPGA board, demonstrated remarkable accuracy in distinguishing between normalpatterns and those indicative of schizophrenia. This achievement with pre-processing offMRI images and underscores the potential of AI-driven tools in assisting healthcareprofessionals with early diagnosis.

[0068] Real-Time Processing: The FPGA based hardware acceleration capabilitiesfacilitated real-time processing of brain scans, opening doors to rapid and on-the-spotassessments, particularly in clinical settings where time is of the essence.

[0069] Resource Efficiency: The FPGA based implementation demonstratedresource-efficient utilization, highlighting the scalability and cost-effectiveness of FPGAbasedsolutions for medical applications.

[0070] While this invention has been described in connection with what is presentlyconsidered to be the most practical and preferred embodiment, it is to be understood thatthe invention is not limited to the disclosed embodiments, but, on the contrary, is intendedto cover various modifications and equivalent arrangements included within the scope ofthe appended claims.

Claims

1. A device for automatic schizophrenia disease detection in a subject patient using functional Magnetic Resonance Imaging (fMRI) images, comprising: plurality of field programmable gate array (FPGA) using convolution neural network (CNN) and a processor configured to execute instructions to perform operations, comprising: performing data analysis on a data comprising plurality of fMRI images of the subject using a Statistical Parameter Mapping (SPM), wherein plurality of said data is pre-processed with the SPM and a Gaussian filter for realigning one or more individual volumes together; estimating the value of each individual volume which is out of alignment with respect to a reference volume, wherein the reference volume is set in a field which allows specifying whether the one or more volumes aligned to a mean of all the volumes; analyzing one or more types of data including a realignment data, a rotation data, and a translation data for measuring change in the movement of the subject, wherein if the subject moved beyond the size of a voxel, or a sudden spike of more than half of the voxel size is detected, it is removed for further processing; aligning a functional data to an anatomical data, the functional data comprises one or more reference images and the anatomical data comprises one or more source images, wherein the reference image remains stationary and the source image is arranged to move around until a best fit is found between the reference image and the source image; examining the subject with mean functional and anatomical images to identify the alignment in ventricles thereof; segmenting the anatomical images in order to align the difference tissue types to illustrate the activation of each element and normalizing all data to standardized space and further aligning brains outlines and internal structures of the subject; and smoothing the data by combining one or more signals and canceling noise by using spatial smoothing with Gaussian kernel Intensity technique.

2. The device as claimed in claim 1, wherein the realignment data provides movement of the subject in a three dimensional coordinate axes space, the rotation data indicates rotation of the subject around each coordinate axis, and the translation data provides movement of the subject along each coordinate axis.

3. The device as claimed in claim 1, wherein the CNN-based hardware automatically detects and classifies fMRI images.

4. The device as claimed in claim 1, wherein the pre-processing of fMRI images utilises the SPM technique along with Gaussian filter, and the SPM is a SPM12.

5. The device as claimed in claim 1, wherein the FPGA-based hardware is reconfigurable.

6. The device as claimed in claim 1, wherein a Xilinx FPGA ZC706 valuation board is used to synthesize hardware architecture to compare with the CNN based software to find accuracy.

7. The device as claimed in claim 1, wherein the FPGA is connected using a serial interconnect and a universal asynchronous receiver-transmitter port.

8. The device as claimed in claim 1, the subject data is stored in an on-chip memory of the FPGA, as well as in a secondary memory for further analysis.

9. A method for automatically detecting schizophrenia disease in a subject patient using functional Magnetic Resonance Imaging (fMRI) images, comprising: performing data analysis on a data comprising plurality of fMRI images of the subject using a Statistical Parameter Mapping (SPM) by pre-processing individual subject data, wherein pre-processing includes motion correction, normalization, filtering, slice timing correction, artefact removal, and Gaussian filter smoothing of said data; identifying differences between a Normal group and a SZ group by evaluating data from both the groups using one or more statistical models, and the data is interpreted by visualizing and reporting group-level brain activity variations between the Normal and SZ groups; estimating the deviation between each individual volume and a reference volume, wherein a parameter of the reference volume specifies whether the volumes align to the mean of all volumes or to the first volume; analyzing realignment of subject movement in x, y, and z dimensions, wherein a rotation of the subject is measured around each axis, and a translation of the subject movement is measured along each axis; eliminating a subject from further analysis if the subject went beyond a voxel or a sudden spike larger than half of a voxel is measured; aligning a functional data to an anatomical data using two fields: a first field includes one or more reference images and a second field includes one or more source images, wherein the functional is used as a reference and the anatomical data is used as source images such that the reference image remains stable while the source image is moved to get the greatest match; examining the subject with mean functional and anatomical images; using anatomical picture segmentation to align different tissue types, wherein the segmentation shows how each element is activated, and standardize the data and further align brain contours and architecture; and smoothing the data by averaging one or more signals and cancelling noise using a Gaussian kernel by increasing the data intensity.

10. The method as claimed in claim 9, wherein Function connection, SBC, ROI, and second-level analysis are conducted before providing to the FPGA based CNN.