WO2023281317A1

WO2023281317A1 - Method and system for analyzing magnetic resonance images

Info

Publication number: WO2023281317A1
Application number: PCT/IB2022/050645
Authority: WO
Inventors: Siva Raja P.M.; Ramanan K
Original assignee: P M Siva Raja; Ramanan K
Priority date: 2021-07-06
Filing date: 2022-01-25
Publication date: 2023-01-12

Abstract

Embodiments of the present disclosure provide a system for analyzing Magnetic Resonance Images of the subject from at least one MR image associated with a cranium of said subject. The system includes an image processing unit (200) to obtain a plurality of MR images from at least one MR image source, generate at least one de-noised MR image by extracting a noise from said at least one MR image, segment said at least one de-noised MR image using BFC for identifying a core and edema regions in said at least one de-noised MR image and extract features from said core and edema regions of said at least one segmented de-noised MR image.

Description

METHOD AND SYSTEM FOR ANALYZING MAGNETIC RESONANCE IMAGES

TECHNICAL FIELD

[0001 ] The presently disclosed subject matter generally relates to a field of medical systems, and more particularly, the present subject matter relates to a method and a system for analyzing Magnetic Resonance Images.

BACKGROUND

[0002] In general, Magnetic Resonance Image (MRI) carries out an important role to study regarding human brain. Through MR images, the valuable information concerning structure of soft tissue can be evaluated. For the manual interpretation, the quantity of data is to an extreme degree unreasonably so, the need for the computerized image analysis tools is highly increased. For the location identification and brain tumour growth imaging, the MR images is more often utilized because of the greater determination.

[0003] Further, the exact segmentation of the MR images is significant for identifying the brain tumour by computer assisted clinical tool. The brain image segmentation is needed for the identification of the brain tumour. The manual brain MRI segmentation is a challenging assignment. Moreover, for this, non-uniform segmentation, lots of time, non-repeatable assignment is needed as well as segmentation outcomes. Sometimes, it differs from specialist to specialist thus the computer assisted tool is highly supportive in this situation. Additionally, the obtained features of MRI image are considered as the crucial part since it indicates image in its compact shape. Some other methods are also used for the feature extraction that allows the classifiers to categorize the tumour either as normal or abnormal.

[0004] In general, the brain tumour segmentation as well as categorization differentiates the malignant and benign cells in the brain and categorizes the malignant cells for detecting its level. The existing techniques are time-consuming as well as insufficient in making decisions.

[0005] Thus, there exists is a need of a mechanism providing a computerized categorization of the brain tumour to overcome aforementioned issues (i.e., time-consuming, inaccurate data, or the like) experienced by said advanced approaches for carrying the brain tumour segmentation as well as categorization.

[0006] Also, it is an object of the present invention to overcome or ameliorate the above discussed disadvantages of the prior art, or at least offer a useful alternative.

SUMMARY

[0007] To overcome the above-mentioned limitations and problems, the present disclosure provides a method and a system for analyzing Magnetic Resonance Images of a subject. [0008] An embodiment of the present disclosure provides a system or an apparatus for analyzing Magnetic Resonance Images of the subject from at least one Magnetic Resonance (MR) image associated with a cranium of said subject. The system is implemented using machine learning. The system includes an image processing unit coupled to a tumour classification unit. The image processing unit is configured to obtain a plurality of MR images from at least one MR image source, generate at least one de-noised MR image by extracting a noise from said at least one MR image, segment said at least one de-noised MR image using Bayesian fuzzy clustering (BFC) for identifying a core and edema regions in said at least one de- noised MR image and extract features from said core and edema regions of said at least one segmented de-noised MR image using an image feature extraction mechanism comprising information theoretic measures, wavelet packet Tsallis entropy and scattering transform. Further, the tumour classification unit can be configured to classify said brain tumour based on said at least one extracted feature using deep auto encoder (DAE) and Softmax regression. The classification of said brain tumour is executed by obtaining first identifiers associated with a plurality of medical images from at least one information repository system, categorizing said at least one extracted feature of said at least one segmented de-noised MR image based on said first identifiers using said DAE and Softmax regression, and classifying said brain tumour based on said categorization.

[0009] In an embodiment, the features extracted from said core and edema regions of said at least one segmented de-noised MR image using said information theoretic measures comprises entropy, mean, variance, kurtosis and skewness corresponding to said at least one segmented least de-noised MR image of a pre-defined modality.

[0010] In an embodiment, the features extracted from said core and edema regions of said at least one segmented de-noised MR image using said scattering transform corresponds to texture of said at least one segmented de-noised MR image.

[0011] In an embodiment, the first identifiers indicate pre-defined labels or information associated with said plurality of medical images.

[0012] In an embodiment, said tumour classification unit is further configured to update said at least one information repository system with second identifiers, where said second identifiers is associated with said categorization of at least one extracted feature of said at least one segmented de-noised MR image.

[0013] In an embodiment, said second identifiers indicates pre-defined labels or information associated with said at least one segmented de-noised MR image. [0014] Another embodiment of the present disclosure provides a method for analyzing Magnetic Resonance Images of the subject from at least one Magnetic Resonance (MR) image associated with a cranium of said subject. The method is implemented using machine learning. The method includes obtaining a plurality of MR images from at least one MR image source, generating at least one de-noised MR image by extracting a noise from said at least one MR image, segmenting said at least one de-noised MR image using Bayesian fuzzy clustering (BFC) for identifying a core and edema regions in said at least one de-noised MR image, and extracting features from said core and edema regions of said at least one segmented de-noised MR image using an image feature extraction mechanism comprising information theoretic measures, wavelet packet Tsallis entropy and scattering transform. Further, the method includes classifying said brain tumour based on said at least one extracted feature using deep auto encoder (DAE) and Softmax regression. The method for classifying said brain tumour includes obtaining first identifiers associated with a plurality of medical images from at least one information repository system, categorizing said at least one extracted feature of said at least one segmented de-noised MR image based on said first identifiers using said DAE and Softmax regression, and classifying said brain tumour based on said categorization.

[0015] Other and further aspects and features of the disclosure will be evident from reading the following detailed description of the embodiments, which are intended to illustrate, not limit, the present disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The illustrated embodiments of the disclosed subject matter will be best understood by reference to the drawings, wherein like parts are designated by like numerals throughout. The following description is intended only by way of example, and simply illustrates certain selected embodiments of devices, systems, and processes that are consistent with the disclosed subject matter as claimed herein.

[0017] Figure 1 is a schematic diagram illustrating a system for analyzing Magnetic Resonance Images, in accordance with an embodiment of the present disclosure;

[0018] Figure 2A is a block diagram illustrating various elements of an image processing unit, in accordance with an embodiment of the present disclosure;

[0019] Figure 2B is a Wavelet packet Tsallis entropy, in accordance with an embodiment of the present disclosure;

[0020] Figure 3A is a block diagram illustrating various elements of a tumour classification unit, in accordance with an embodiment of the present disclosure;

[0021] Figure 3B illustrates Deep autoencoder and Softmax regression integration, in accordance with an embodiment of the present disclosure;

[0022] Figure 4 is a flowchart diagram illustrating a method for analyzing Magnetic Resonance Images, in accordance with an embodiment of the present disclosure; [0023] Figures 5a-5c illustrates a Brain MRI outputs, in accordance with an embodiment of the present disclosure;

[0024] Figure 6 is a process flow diagram for analyzing Magnetic Resonance Images, in accordance with an embodiment of the present disclosure; and

[0025] Figures 7-15 are graphs analysis depicting a comparison between proposed method of present invention and existing/conventional methods, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION

[0026] The following detailed description is made with reference to the figures. Exemplary embodiments are described to illustrate the disclosure, not to limit its scope, which is defined by the claims. Those of ordinary skill in the art will recognize a number of equivalent variations in the description that follows.

[0027] The functional units described in this specification have been labeled as devices. A device may be implemented in programmable hardware devices such as processors, digital signal processors, central processing units, field programmable gate arrays, programmable array logic, programmable logic devices, cloud processing systems, or the like. Nevertheless, the executable of an identified device need not be physically located together, but may comprise disparate instructions stored in different locations which, when joined logically together, comprise the device and achieve the stated purpose of the device. Reference throughout this specification to “a select embodiment,” “one embodiment,” or “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Thus, appearances of the phrases “a select embodiment,” “in one embodiment,” or “in an embodiment” in various places throughout this specification are not necessarily referring to the same embodiment. Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided, to provide a thorough understanding of embodiments of the disclosed subject matter. One skilled in the relevant art will recognize, however, that the disclosed subject matter can be practiced without one or more of the specific details, or with other methods, components, materials, etc. In other instances, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring aspects of the disclosed subject matter.

[0028] In bio medical images, tumour detection is remedied as a significant method in the latest growth of clinical investigation, which are achieved by clinical specialists. In the beginning, imaging processes are utilized to document medical pictures for identifying the tumour. The segmentation is a compulsory step in the medical imaging that should be executed with accurateness however it consumes enormous time and experience. This difficult assignment as well as timewasting makes radiologist to move towards the requirement for the semi automatic segmentation technique. Also, that shows some possibilities of reducing the disadvantages of computerized segmentation technique at the same time the radiologists can also manage the segmentation method. The infectious area is separated using the image segmentation from the remaining part of the picture. The treatment planning is supported while the precise segmentation technique assists to decide the location and size of the tumour. An expert physician has to set both the early settings and information about training for the categorization. For the identification of various kinds of tumour, so many investigations have been executed on the basis of the extraction of the visual data from the medical images.

[0029] Thus, the brain tumour segmentation as well as categorization is the developing investigation region to differentiate the malignant and benign cells in the brain and to categorize the malignant cells for detecting its level. The advanced methods shortage the computerized categorization and more time-consuming as well as insufficient in making decisions.

[0030] Unlike conventional methods and systems, the proposed method and system provides a computerized categorization of brain tumour to overcome the issues faced by the advanced approaches utilizing the optimization mechanism to train the Softmax regression classifier and hybrid deep auto encoder.

[0031] The proposed system and method can be used to enhance the tumour grade identification process from the MRI by the assistance from the deep learning and bio-inspired computing mechanisms.

[0032] Figure 1 is a schematic diagram illustrating a system 1000 for automatically classifying said brain tumour, in accordance with an embodiment of the present disclosure.

[0033] The system 1000 includes an image processing unit 200 configured to process an image 100. The image 100 may include any medical image that can aid in identifying and diagnosing the brain tumour of a subject (e.g., living beings). For example, the image 100 can be an MRI obtained from MRI scanning machines. In alternative embodiments (or optionally), the image 100 may also be an X-ray image obtained from X-ray scanning machines or CT image obtained from CT scanning machines.

[0034] In brain tissue, evolution of abnormal cells constitutes the brain tumor. It can be produced from the cells that have spread to the brain from a cancer. It can arise from abnormal growth of the cells inside the brain. The two kinds (or types) of tumours are cancerous and non- cancerous. It has proven that a third level of brain tumours are also cancerous which will end with death as like as the highest point of the cancer. The early identification as well as treatment are significant for subject (i.e., patient(s) being diagnosed with tumour). Both benign and malignant are also considered as the serious brain tumour. Finally, the tumour increases in the head which can flatten the other growth. The primary tumour and the secondary tumour are the two choices of the brain tumour. The primary tumour starts in the brain tissues and moreover the secondary tumour extends to the skull from other part of the body. [0035] In an embodiment, for the detection of the tumours, the MR image 100 is utilized and provides the information about an organization of human soft tissue that is involved with the radiology to observe the arrangement as well as the behavior of the human body. As well as the MR image 100 is more significant and helpful in the medicinal imaging because it offers various change amid the disparate soft tissues of the body. Also, the MR image 100 pass through several information together with the tumour. From numerous techniques, the MR image 100 is identified because of its excellence.

[0036] Unlike to conventional mechanism, as described above, the present invention aims at analyzing and enhancing the tumour grade identification process from the MR image 100 using machine learning mechanism such as deep learning and bio-inspired computing mechanisms. Further, the present invention utilizes Bayesian fuzzy clustering for segmentation of the MR image 100 and deep autoencoder with Softmax regression to categorize tumours from the MR image 100.

[0037] The MR image 100 can be directly obtained from an MR image source (not shown in FIGs) or can be externally provided (for example, copy of stored/parked MR images) to the image processing unit 200. The MR image source can be the MRI scanner or any other scanner capable to producing the MR image 100.

[0038] In an embodiment, the image processing unit 200 can be configured to process the MR image 100 in order to upskill a quality of the MR image 100 and to identify key features from the MR image 100 that can aid in accurate detection and classification of brain tumours. For example, the image processing unit 200 can be configured to extract a noise from the MR image 100 and generate a de-noised MR image. Further, the image processing unit 200 can be configured to segment said at least one de-noised MR image using Bayesian fuzzy clustering (BFC) for identifying a core and edema regions in said de-noised MR image. Furthermore, the image processing unit 200 can be configured to extract features from said core and edema regions of said segmented de-noised MR image using an image feature extraction mechanism comprising information theoretic measures, wavelet packet Tsallis entropy and scattering transform (as detailed below in Figure 2).

[0039] The system 1000 further incudes a tumour classification unit 300 for classifying the brain tumour from the MR image 100 (or equivalent de-noised MR image or segmented de- noised MR image) produced by said image processing unit 200. The classification of the brain tumour, according to present invention, is executed by using a Softmax regression classifier and Hybrid Deep auto encoder (DAE). The proposed method of classification can aid with an information related to a stage of the brain tumour (or different levels of tumour) using the extracted features of the segments yielded based on said BFC. [0040] The classification of said brain tumour is by categorizing said extracted features using the DAE and Softmax regression. The tumour classification unit 300 can be configured to obtain first identifiers associated with a plurality of medical images from an information repository system 400. In an embodiment, the information repository system 400 can include, for example, a knowledge -based dataset/database such as multimodal Brain Tumor Segmentation (BRATS). The BRATS contains different low- and high-grade gliomas brain images as well as the mixed high- and low-grade gliomas brain images for testing purpose. The tumour classification unit 300 utilizes a skull removed images. Thus, by inserting the results of segmentation of the MR image 100, the performance of few statistics values is esteemed to the presented BRATS system. The first identifiers indicate a pre-defined labels or information (such as type of tumour, level of tumour, type of cancer or any other information relevant to identify and classify the tumours) associated with said plurality of medical images (such as for example, MR images).

[0041] The system 1000 can access the information repository system 400 remotely using a server or any cloud-based network.

[0042] The system 1000 may be implemented using any electronic device or apparatus (not shown in Figs) capable to executing operations of the image processing unit 200 and said tumour classification unit 300.

[0043] Figure 2 is a block diagram illustrating various elements of the image processing unit 200, in accordance with an embodiment of the present disclosure.

[0044] The image processing unit 200 includes a de-nosing engine 110, a segmentation engine 120, a feature extraction engine 130, and a memory 140.

[0045] The de-nosing engine 110 can be configured to generate the de -noised MR image 100 by extracting the noise from the MR image 100. The MRI image(s) 100 have several intensity ranges that are affected by bias fields differently. So adopted a robust intensity normalization technique to make MRI scans of different patients, and also modifying the bias field of MRI data. The compression of tumour detection depends on the regions of interest which are typically noisy nature and low contrast in the growth of MRI image compression scheme. Hence, an image denoising and development may be compulsory to preserve the image quality, highlighting image features and suppressing the noise. There are lots of filters which have been used for filtering the images, some of them corrupt the minute information of the image and some conventional filters will process the image smoothing and subsequently harden the edges of the image. The de-nosing engine 110 is responsible for producing better clarity of the MR image 100 by eliminating the noise from said MR image 100. [0046] The de-nosing engine 110 can be configured to remove noise from the MR image 100 using non-local mean filter. It does not update a pixel’s value with an average of the pixels around it, rather than updates a weighted average of the pixels judged to be most attractive. The weight of each pixel depends on the distance between its intensity grey level vector and that of the target pixel. De-noised image of each pixel of the non-local means is calculated as,

where, j is the noisy image, de-noised image is denoted by N and weights w(i,j) meet the following conditions 0 < rv(z, _j) < 1. Weighted average of all the pixels is based on the comparison between the neighborhoods of pixels i and j.

[0047] The segmentation engine 120 can be configured to utilize the Bayesian fuzzy clustering (BFC) for identifying the core and the edema tumour regions from the MR image 100. It yields the segments for extracting features. The Bayesian fuzzy clustering (BFC) forms a Bayesian model that is a joint probability of data points and parameters such as Gaussian prior distribution, fuzzy cluster prior, FDL (fuzzy data likelihood) or data likelihood distribution. Based on the fuzzy data likelihood, the BFC model is explained in terms of data matrix, fuzzy membership, cluster prototypes, number of data points and number of data points in the cluster.

[0048] The probability of data is directly proportional to the product of the cluster prototypes and normal probabilities is known as FDL. Among the data points, the groups are formed they are share the mean values which are termed as the cluster prototypes.

where,

X (c_h , m, C) denotes the normalization constants

C indicates the cluster E is the total number of the clusters The cluster number is denoted by k m is the fuzzifier

[0049] By using the Bayesian model, FCP is utilized for the replication of the FCM performance.

[0050] The FCP contains three factors, such as X c_h,m,C) ,

and Dirichlet c_h I _Li) , where, c_hk signifies the membership of the data point d_h in the cluster k, m identify the mean value.

[0051] The initial factor discards the normalization constant of the FDL. For providing the higher member-ship values, the second factor is used. The positivity is represented by the third factor and it provide the additional flexibility and capability to the clustering mechanism. The Gaussian prior distribution is defined in terms of mean of the dataset, data covariance, user set parameter that affects the strength. The joint probability of the data and parameters are formed by using these all factors.

[0052] The denoising process may include local noise reduction, inter-slice intensity variation correction and intensity inhomogeneity correction. The denoising may be done using filtering techniques, such as mean filtering, Gaussian filtering, linear filtering, anisotropic diffusion filtering or the like. The local noise reduction may be done using mean filtering, where each pixel's intensity value is replaced with the mean of its neighbours. The neighbours may be defined by a square window centered at the pixel. A more popular method of noise reduction is through Gaussian filtering. Alternatively, the denoising may be done using Segment Assimilating Nucleus filter, wherein contribution of neighbouring pixels is weighted through a Gaussian in spatial and intensity domain. The intensity domain smoothens near thin lines and corners by weighing pixels on thin line more heavily. It also implements heuristic to account for impulse noise.

[0053] The inter-slice intensity variation correction reduces sudden intensity variations between adjacent slices. Further, the intensity inhomogeneity correction estimates a 3D (dimensional) inhomogeneity field of same size as volume, which represents the intensity inhomogeneity. The estimated field may be used to generate an image where variance in intensity for each tissue type is reduced, and differences between tissue types are increased.

[0054] The feature extraction engine 130 can be configured to extract robust features from the core and the edema tumour regions of the MR image 100 using information theoretic measures, wavelet packet Tsallis entropy and scattering transform. The (image) feature extraction is processed in three major steps. In the first step, the modality corresponding to the separate segment is attained to produce a set of five features (detailed below using equation 10) using the information theoretic measures. In the second step, the scattering transform is employed for feature extraction and in third step, the wavelet packet Tsallis entropy is used for the extraction of the features.

[0055] According to, the information theoretic measures, the useful information from an original signal is extracted, information that is relevant for the tumour classification and also has a more compact representation, suitable for use in a classifier. This can be achieved simply through extraction, in which elements of the original data vector are kept, or through a transform, which can project the original data in a different, lower-dimensional space.

[0056] In an embodiment, a main image -based features used may be intensities. To take into account neighborhood information at different scales and characterize local image textural properties, the responses of above-mentioned filters of the images as features may directly be employed.

[0057] Further according to, the Wavelet packet Tsallis entropy (WPTE), a combination of discrete wavelet packet transform (DWPT) and Tsallis entropy (TE) is called as Wavelet packet Tsallis entropy. Standard discrete wavelet transforms (DWT) pass only the earlier approximation sub-band (A) via quadratic mirror filters. Nevertheless, the DWPT pass both approximation sub band, and three detail sub-band to the filters. Thus, for an n-level decomposition, DWT provides (3n + 1) sub-bands, while DWPT provides sub-bands, as shown in Figure 2B.

[0058] An improved version of traditional Shannon entropy is known as Tsallis entropy. Assume p represents the probability mass function, h the gray level value of any sub-band coefficient, H the total number of gray level values. The Tsallis entropy E_r announced a new entropic-index parameter q, and its definition is written as

[0059] Parameter q measures the nonextensivity degree of the probability mass function. When q approximates to 1, the Tsallis entropy will degrade to Shannon entropy. Figure 2B shows the diagram of WPTE, where A means approximation, H horizontal, V vertical, and D diagonal. At first, a brain slice image (e.g., the MR image 100) is imported, then a 3 -level discrete wavelet packet transform is carried out on said MR image 100. The coefficients of all 64 sub-bands generated by the 3-level DWPT decomposition are stored in a memory 140. The Tsallis entropy over each sub-band is then calculated. Finally, a 64-element vector for each brain image is provided as an output.

[0060] Furthermore, according to scattering transform, texture features from the MR image 100 are extracted using the scattering transform (ST). The specific segments from the individual modality of the MR image 100 of subject brain is an input to the scattering transform. For classification, the features are extracted which is essential to be invariant with respect to the transformation. The scattering transform construct invariant, stable and informative representations through a unitary transform, non-linear, which delocalizes the info into scattering decomposition paths. They are calculated with a cascade of wavelet modulus operators. It corresponds to a convolutional network where filter coefficients are given by a wavelet operator. [0061]

[0062] In the wavelet modulus coefficients, formation of the scattering coefficients using the successive wavelet modulus transforms. By using the convolution of the wavelet modulus coefficients, the higher order scattering coefficients are achieved with the newly produced wavelets of orientation and different scale. The higher order coefficients produce the robustness of the extracted features. The scattering transform is Fipschitz continuous to deformations. [0063] The feature extraction may be done using an edge implementing technique on the MR image, where a filter such as Gabor filter may be utilized to obtain a first set of edges. The MR image may be transmitted into a gray-value image “I.” Different sizes of Gabor filters may be applied as the convolution kernels to process the gray-value image I.

[0064] Further, an edge pooling technique may be performed on the first set of edges in order to obtain a second set of edges. The edge pooling technique may discard noisy and redundant edges from the first set of edges resulting in reduced number of the second set of edges. The edge pooling technique may utilize a MAX operation to obtain the second set of edges from the first set of edges.

[0065] The MR image may be sampled to obtain a set of patches like a set of edge patches, a set of color patches and a set of texture patches. Typically, a “patch” is a region of pixels sampled at a specific position of the MR image in multiple resolutions and at multiple orientations. The set of patches from the set of edge patches, the set of color patches and the set of texture patches may be selected. Any sampling technique may be utilized obtain the set of edge patches, color patches and texture patches, respectively. For example, edge patches may be pixels that are sampled from the second set of edges, color patches may be sampled at various resolutions and may include color information features such as color, color histograms, color means and variances, and shape characteristics such as elongation and spread. Further, texture patches may be sampled and include texture information such as scale, structure, shape and orientation characteristics.

[0066] Each patch (excluding the first) may be selected by determining the patch that: ( 1 ) is not already in the set of patches, and (2) is the “farthest” from the immediately preceding patch according to Euclidean distance.

[0067] Further, a part selection technique may be performed on the set of patches to obtain a first set of parts. Each of the first set of parts may describe larger regions of the MR image than the edge patches, color patches and texture patches. In some embodiments, the first set of parts may be obtained by examining the selected patches in each of the sets of patches obtained from a plurality of MR images that are used to train the technique. If the selected patches occur frequently enough within sets of patches, these patches can be selected as parts.

[0068] Additionally, a part pooling mechanism may be performed on the first set of parts to obtain a second set of parts. The part pooling may be performed to discard noisy and redundant parts from the first set of parts. Similar to the aforesaid edge pooling technique, the part pooling mechanism may utilize a MAX operation to obtain the second set of parts from the first set of parts. Lastly, the second set of parts may be mapped to identified area in the MR image. The mapping may be based on the degree of matching of the second set of parts with sampled parts in the plurality of MR images utilized to train the technique. [0069] Figure 3 is a block diagram illustrating various elements of the tumour classification unit 300, in accordance with an embodiment of the present disclosure. The tumour classification unit 300 includes, for example, a categorization engine 210 comprising a coding engine 202, a regression engine 220. Further, the tumour classification unit 300 includes a memory 230 configured to store type of brain tumours that are classified using said tumour classification unit 300.

[0070] In an embodiment, the extracted robust features (obtained using the feature extraction engine 130) are used as the input features to the tumour classification unit 300 for classifying the brain MRI based on class labels. The whole network weights are fine-tuned to reach the global optimum through optimization mechanism/instructions. The classification is performed by obtaining the first identifiers associated with the plurality of medical images from the information repository system (400). Further, the categorization engine 210 can be configured to categorize said at least one extracted feature of said at least one segmented de-noised MR image based on said first identifiers using said coding engine 210 and regression engine 204.

[0071 ] The coding engine 202 can include a deep auto-encoder (DAE) also known as stacked auto-encoder and it is a primitive deep network, which contains auto-encoder with multiple hidden layers. It has greater expressive power. For classification problem, the Softmax classifier is commonly nominated as the output layer. Auto-encoder is an unsupervised neural network which used to encode input samples into some representations. So that the inputs can be reconstructed from that representations with minimum reconstruction error.

[0072] With a linear transformation and a sigmoid function, the encoding and decoding functions are comprehended by a composite function. The pre-training trains the parameters of each layer individually to produce better results. Fine-tuning is a global optimization strategy which is used in neural network. It can enhance the performance of the DAE. Fine-tuning tries to diminish the difference between the true label and the output. The squares error cost respect to a single sample.

[0073] The regression engine 204 can include the Softmax classifier as the last layer in the deep neural network. The aim of Softmax classifier is classifying the learned features from deep auto-encoders. The DAE can form the deep network structure through the multilayer stack. It can be used for feature learning. However, this auto-encoder has no ability to classify. Therefore, the present invention provides a deep neural network structure that combines the deep auto-encoder (or said coding engine 202) and Softmax regression (or said regression engine 204). In multiple classifications, the Softmax regression is an extension of the logistic regression model. The category tag of the logistic regression can only take two values, whereas the Softmax tag can take on multiple values. Figure 3c shows the diagram of deep autoencoder and Softmax regression.

[0074] All the parameters have been initialized in the pre-training step for deep learning. Hence, the parameters can be updated by minimizing the energy function using stochastic gradient descent to finish DAE fine-tuning. Aforementioned optimizing instructions can be used for problems either constrained or unconstrained. The outcome obtained from the problem reaches the optimal solution, by preventing the worst solution. The proposed method makes use of objective function to extract the abnormal portion from the MR image 100.

[0075] Figure 4 is a flowchart (400) illustrating a method for analyzing Magnetic Resonance Images, in accordance with an embodiment of the present disclosure. The operations (402-406) are performed by the image processing unit 200 and the operations (408-410) are performed by the tumour classification unit 300.

[0076] At step 402, the method obtaining the plurality of MR images 100 from at least one MR image source. At step 404, the method includes generating the at least one de-noised MR image by extracting the noise from said at least one MR image 100. At step 406, the method includes segmenting said at least one de-noised MR image using the Bayesian fuzzy clustering (BFC) for identifying the core and edema regions in said at least one de-noised MR image. At step 408, the method includes extracting features from said core and edema regions of said at least one segmented de-noised MR image using the image feature extraction mechanism comprising said information theoretic measures, the wavelet packet Tsallis entropy and the scattering transform. Further, at step 410, the method includes classifying said brain tumour based on said at least one extracted feature using a deep auto encoder (DAE) and Softmax regression.

[0077] The various actions, acts, blocks, steps, or the like in the flow chart (400) may be performed in the order presented, in a different order or simultaneously. Further, in some embodiments, some of the actions, acts, blocks, steps, or the like may be omitted, added, modified, skipped, or the like without departing from the scope of the invention.

[0078] Following sections corresponds to experimental analysis and performance analysis.

[0079] Performance metrics: To measure the precision, specificity, F-measure, sensitivity, false positive and negative rate, balance error rate, negative predictive value and accuracy, the confusion matrix of true and false positive then true and false negative were used.

True positive (TP): The analysis result is positive in the occupancy of the medical abnormality.

True negative (TN): In the lack of the clinical abnormality, the classification outcome is negative. False positive (FP): In the absence of the medical abnormality, the categorization result is positive.

False negative (FN): In the presence of the clinical abnormality, the classification result is negative.

Accuracy: The amount of the perfection of the recognition.

Sensitivity: The sensitivity or the true positive rate (TPR) is defined as the sum of positives predictable correctly.

Specificity: The specificity or the true negative rate (TNR) is defined as the total of negatives acknowledged suitably.

Precision: The proportion of predicted positives which are genuine positive.

F-measure: The harmonic mean of precision and sensitivity.

False positive rate (FPR): The fraction of all negatives still provides the positive test outcomes that implies the contingent likelihood of a positive test outcome indicated an occasion that was absent.

False negative rate (FNR): The ratio of positives which convey negative test results with the test, that involves the contingent likelihood of a negative test result given that the condition being watched for is available.

Balanced error rate (BER): The average of the error rate on the positive class and the error rate on the negative class.

Negative predictive value (NPV): Negative predictive value is the probability that subjects with a negative screening test really don't have the illness.

[0080] Performance Analysis: The deep auto encoder using the BRATS database and dissimilar classifiers of existing classifiers for categorization of brain tumours such as, for example, Deep neural network (DNN), Artificial neural network (ANN), K-Nearest Neighbour (KNN) and Multi-SVM using the comparable datasets are shown in Figures 7-15. The performance is valued by relating its classification outcomes with traditional classifier system which uses the DAE and Softmax regression-based tumour classification framework. The results are displayed in the accompanying graphs. The test results presented that the classification accuracy, sensitivity, specificity, precision, F-measure, false positive and negative rate, balanced error rate and negative predictive values are accomplished through the deep auto encoder and Softmax regression is better than the results attained through other classifiers. Figure 5 demonstrates the brain MRI output of the different procedure used in the recommended system. Figures 5(a) and 5(b) shows the original input and noise removed of brain image individually. The deep auto encoder and Softmax regression is illustrated in Figure 5(c).

[0081] Unlike to conventional system and method, the proposed mechanism provides an accurate brain tumour classification using hybrid deep auto encoder with Bayesian fuzzy clustering segmentation scheme. The method of Bayesian fuzzy clustering is utilized which provides the positive regions from the MR image 100. For detecting the presence of edema and non-tumours, the vigorous features are gotten from the effective regions that understandable the occupancy of the abnormalities. Information theoretic measures, wavelet packet Tsallis entropy and scattering transform is utilized for feature extraction process. To categorize the brain MR images 100, the deep auto encoder and Softmax regression is applied for classification by using the information repository system 400. The proposed system has the highest accuracy (98.5%) when compared to other existing systems.

[0082] Figure 6 illustrates a process flow (602-610) for classifying brain tumour. The process flow of proposed brain tumour classification approach consists of four main phases: pre processing 602a-b, segmenting image slices using Bayesian fuzzy clustering 604a-b, robust feature extraction 606a-b by using dataset of tested MR images 608 and tumour classification 610. The tumour may be, for example, classified as normal and/or abnormal.

[0083] Figure 7 shows the performance of accuracy of proposed method of present invention and the existing approaches of the DNN, ANN, KNN and Multi-SVM. The proposed method attains the high accuracy of 98.5% than other existing approaches/systems. The graphs indicate the proposed methodology is achieved the highest performance when associated to other classifiers.

[0084] Figure 8 shows the comparison of sensitivity of proposed and existing methods. The graph produce sensitivity more than other classifiers using the deep auto encoder and Softmax regression, the sensitivity is attained for the proposed approach and it is compared to other classifiers. The proposed approach obtained the high sensitivity than other approaches.

[0085] Figure 9 illustrates the specificity of proposed method and existing classifiers of DNN, KNN, ANN and Multi-SVM. The proposed method achieved the high specificity when compared to other classifiers. The proposed method gives the better results than other existing techniques. The proposed deep auto encoder and Softmax regression is given the better performance for brain tumour classification.

[0086] Figure 10 displays the performance of precision of proposed and existing classifiers. The graph represents the proposed approach incorporating said deep auto encoder and Softmax regression achieves the highest precision when compared to other classifiers of DNN, KNN, ANN and Multi-SVM. The proposed method aims at providing the better performance than existing approaches.

[0087] Figure 11 shows the performance analysis of F-measures of proposed and existing method. In this graph, the proposed provides the better results than other conventional or existing methods. The proposed method incorporating the deep auto encoder provides the better F- measures when compared to other methods of DNN, KNN, ANN, multi-SVM. [0088] Figure 12 illustrate the false positive rate comparison of our proposed and existing methods. The graph displays the proposed obtained the less false positive rate than the other classifiers of KNN, ANN, DNN and Multi-SVM. The proposed method provides the better results when compared to other existing segmentation approaches.

[0089] Figure 13 illustrate the performance of false negative rate of proposed of DAE and existing of Multi-SVM, KNN, ANN and DNN. The graph indicates that the proposed approach achieves the less false negative rate than other existing/conventional methods. The proposed method produces the better outcomes for brain tumour classifications.

[0090] Figure 14 provides the BER performance of our proposed of DAE and other classifiers of Multi-SVM, ANN, KNN and DNN. The graph produces the proposed approach of DAE and Softmax regression is obtained less error than other existing classifiers. Our proposed give the better results when compared to other approaches.

[0091] Figure 15 shows the negative predictive value of proposed DAE and existing of DNN, KNN, ANN and Multi-SVM approaches. By compared to others, the proposed method provides the more negative predictive value. The graph clearly indicates the proposed method accomplished the high predictive value for brain tumour classifications.

[0092] For the recommended system, the average of every single factual measure and additionally the numerical measures of other classification techniques are estimated. The figure 6 to 14 speaks to the comparison graph of the proposed system with the other classification system, from that the accuracy, sensitivity, specificity, precision, F-measure and negative predictive value are sensibly higher than other existing method and systems.

[0093] Hence, the performance investigation shows that proposed method and system provides better results/output than the other existing method/system for classification of said brain tumour. The comparison performances such as accuracy, sensitivity, specificity, precision, F- measure, FPR, FNR, BER and NPV likewise similarly superior to alternate strategies. It will be understood that the devices and the databases referred to in the previous sections are not necessarily utilized together method or system of the embodiments. Rather, these devices are merely exemplary of the various devices that may be implemented within a computing device or the server device, and can be implemented in exemplary another device, and other devices as appropriate, that can communicate via a network to the exemplary server device, It will be appreciated that several of the above-disclosed and other features and functions, or alternatives thereof, may be desirably combined into many other different systems or applications. Various presently unforeseen or unanticipated alternatives, modifications, variations, or improvements therein may be subsequently made by those skilled in the art, which are also intended to be encompassed by the following claims.

Claims

We Claim:

1. A system (1000) for analyzing Magnetic Resonance Images of a subject, using machine learning, from at least one Magnetic Resonance (MR) image associated with a cranium of said subject, the system comprising: an image processing unit (200) configured to: obtain a plurality of MR images from at least one MR image source; generate at least one de-noised MR image by extracting a noise from said at least one MR image; segment said at least one de-noised MR image using Bayesian fuzzy clustering (BFC) for identifying a core and edema regions in said at least one de-noised MR image; extract features from said core and edema regions of said at least one segmented de-noised MR image using an image feature extraction mechanism comprising information theoretic measures, wavelet packet Tsallis entropy and scattering transform; and a tumour classification unit (300), communicatively coupled to the image processing unit, configured to classify said brain tumour based on said at least one extracted feature using a deep auto encoder (DAE) and Softmax regression, wherein said tumour classification unit configured to classify said brain tumour by: obtaining first identifiers associated with a plurality of medical images from at least one information repository system (400), categorizing said at least one extracted feature of said at least one segmented de- noised MR image based on said first identifiers using said DAE and Softmax regression, and classifying said brain tumour based on said categorization, and wherein said features extracted from said core and edema regions of said at least one segmented de-noised MR image using said information theoretic measures comprises entropy, mean, variance, kurtosis and skewness corresponding to said at least one segmented least de-noised MR image of a pre-defined modality and wherein said features extracted from said core and edema regions of said at least one segmented de-noised MR image using said scattering transform corresponds to texture of said at least one segmented de-noised MR image.

2. The system (1000) as claimed in claim 1, wherein said first identifiers indicates pre defined labels or information associated with said plurality of medical images, and wherein said tumour classification unit is further configured to update said at least one information repository system with second identifiers, wherein said second identifiers is associated with said categorization of at least one extracted feature of said at least one segmented de-noised MR image, and wherein said second identifiers indicates pre-defined labels or information associated with said at least one segmented de-noised MR image.

3. The system (1000) as claimed in claim 1, wherein the image processing unit (200) comprises a de-nosing engine (110), wherein the de-nosing engine (110) is configured to remove noise from the MR image using a non-local mean filter by updating a weighted average of pixels judged to be most attractive, wherein weight of each pixel depends on distance between its intensity grey level vector and that of a target pixel, and wherein the image processing unit (200) comprises a segmentation engine (120) that utilize the Bayesian fuzzy clustering (BFC) for identifying the core and the edema tumour regions from the MR image (100), wherein the Bayesian fuzzy clustering (BFC) forms a Bayesian model that is a joint probability of data points and parameters such as Gaussian prior distribution, fuzzy cluster prior and FDL (fuzzy data likelihood) or data likelihood distribution.

4. The system (1000) as claimed in claim 1, wherein the image processing unit (200) comprises a feature extraction engine (130) for the image feature extraction, wherein modality corresponding to a separate segment is attained to produce a set of five features - entropy, mean, variance, kurtosis and skewness, using the information theoretic measures, then the scattering transform is employed for the image feature extraction and then, the wavelet packet Tsallis entropy is used for the extraction of the features.

5. The system as claimed in claim 1, wherein feature extraction is done using an edge implementing technique on the MR image, where a filter such as Gabor filter is be utilized to obtain a first set of edges, and wherein an edge pooling technique is performed on the first set of edges in order to obtain a second set of edges, and wherein edge pooling technique discards noisy and redundant edges from the first set of edges resulting in reduced number of the second set of edges.

6. The method as claimed in claim 1, wherein part pooling is performed on a first set of parts to obtain a second set of parts, wherein a part pooling is performed to discard noisy and redundant parts from the first set of parts, and wherein the second set of parts are mapped to identified area in the MR image, wherein the mapping is based on the degree of matching of the second set of parts with sampled parts in the plurality of MR images utilized to train the part pooling.

7. A method, implemented using machine learning, for analyzing Magnetic Resonance Images of a subject from at least one Magnetic Resonance (MR) image associated with a cranium of said subject, the method comprising: obtaining, by an image processing unit (200), a plurality of MR images from at least one MR image source; generating, by the image processing unit (200), at least one de-noised MR image by extracting a noise from said at least one MR image; segmenting, by the image processing unit (200), said at least one de-noised MR image using Bayesian fuzzy clustering (BFC) for identifying a core and edema regions in said at least one de-noised MR image; extract, by the image processing unit (200), features from said core and edema regions of said at least one segmented de-noised MR image using an image feature extraction mechanism comprising information theoretic measures, wavelet packet Tsallis entropy and scattering transform; classifying said brain tumour based on said at least one extracted feature using a deep auto encoder (DAE) and Softmax regression by a tumour classification unit (300) communicatively coupled to the image processing unit; wherein said tumour classification unit configured to classify said brain tumour by: obtaining first identifiers associated with a plurality of medical images from at least one information repository system (400), categorizing said at least one extracted feature of said at least one segmented de- noised MR image based on said first identifiers using said DAE and Softmax regression, and classifying said brain tumour based on said categorization, and wherein said features extracted from said core and edema regions of said at least one segmented de- noised MR image using said information theoretic measures comprises entropy, mean, variance, kurtosis and skewness corresponding to said at least one segmented least de- noised MR image of a pre-defined modality and wherein said features extracted from said core and edema regions of said at least one segmented de-noised MR image using said scattering transform corresponds to texture of said at least one segmented de-noised MR image.

8. The method as claimed in claim 7, wherein said first identifiers indicates pre-defined labels or information associated with said plurality of medical images, wherein said tumour classification unit is further configured to update said at least one information repository system with second identifiers, wherein said second identifiers is associated with said categorization of at least one extracted feature of said at least one segmented de-noised MR image, and wherein said second identifiers indicates pre-defined labels or information associated with said at least one segmented de-noised MR image.

9. The method as claimed in claim 7, wherein the image processing unit (200) comprises a feature extraction engine (130) for the image feature extraction, wherein modality corresponding to a separate segment is attained to produce a set of five features - entropy, mean, variance, kurtosis and skewness, using the information theoretic measures, then the scattering transform is employed for the image feature extraction and then, the wavelet packet Tsallis entropy is used for the extraction of the features.

10. The method as claimed in claim 7, wherein said MR image is sampled to obtain a set of patches comprising a set of edge patches, a set of color patches and a set of texture patches, and wherein edge patches are sampled from the second set of edges, color patches are sampled at various resolutions and include color information features comprising color, color histograms, color means and variances, and shape characteristics such as elongation and spread, and wherein part pooling is performed on a first set of parts to obtain a second set of parts, wherein a part pooling is performed to discard noisy and redundant parts from the first set of parts, and wherein the second set of parts are mapped to identified area in the MR image, wherein the mapping is based on the degree of matching of the second set of parts with sampled parts in the plurality of MR images utilized to train the part pooling.