AU2021106727A4 - Ai system for brain mri/ct malignancy identification and classification using modified cnn and adam optimization. - Google Patents

Abstract

TITLE OF THE INVENTION: "AI SYSTEM FOR BRAIN MRI/CT MALIGNANCY IDENTIFICATION AND CLASSIFICATION USING MODIFIED CNN AND ADAM OPTIMIZATION." ABSTRACT The lifestyle of the people and genetic changes in the body and many other factors cause an increase in neurologic diseases in human that in turn may lead to a life threatening state of human. Due to this scenario, the patients suffering from these neurological diseases struggle to get back to good health. This further points to diagnosis of the exact in the human brain. Even though there have been advancements in the neuro imaging techniques, the proper identification and location of the malignancy and in brain still remains as a challenging task. As to rectify or to reduce the problem, artificial intelligence could play a vital role in classification of images from a set of MRI/CT. Different kinds of methods are normally followed, that has the limitations in accuracy and diagnostic time. Therefore, Artificial Intelligence with Deep learning (DL) is enabled to process the MRI/CT data of a subject. The process of Deep learning verifies the data for classification, the optimize the parameter learning rate and selects the optimization method. In this DL, the modified CNN with Yolo v2 method is chosen to delivery better accuracy in the diagnosis where the marking of the malignancy can be performed in this method. The proposed method uses ADAM (Adaptive Moment estimation) for optimization purposes. The classification can be performed based on the modified CNN transfer learning algorithmic system also, which is capable of searching the random data, and could offer efficiency with acceptable accuracy and with reduced diagnostic time. Page 1 of 1 FIGURE 3: THE PROPOSED SYSTEM MODEL Brain Input Data Data Collection and Sorting MRI/CT Pre-Processing Data f Normalization, Shuffling, Resizng SplitData Training and Validation Modified-ConvolutionalNN Optimization by AGDM Training the layer s and parameters method of CNN's Mod-CNN jgMalignancy Complete the training with marking marking Brain Malignancy Evaluating the Validation Data and Testing

Description

FIGURE 3: THE PROPOSED SYSTEM MODEL

Brain Input Data Data Collection and Sorting MRI/CT

Pre-Processing Data f Normalization, Shuffling, Resizng

SplitData Training and Validation

Modified-ConvolutionalNN

Optimization by AGDM Training the layer s and parameters

method of CNN's

Mod-CNN jgMalignancy Complete the training with marking

marking Brain Malignancy

Evaluating the Validation Data and Testing

EDITORIAL NOTE 2021106727

There are 9 pages of description only.

TITLE OF THE INVENTION: "Al SYSTEM FOR BRAIN MRI/CT MALIGNANCY IDENTIFICATION AND CLASSIFICATION USING MODIFIED CNN AND ADAM OPTIMIZATION."

FIELD OF TIE INVENTION

This invention is intended for the development of a system for quick diagnosis with proper classification to detect the data input image from the data set of Brain MRI/CT using modified CNN, the part of Artificial Intelligence. The modified CNN is used in recognizing objects or images and further process it using different changes in the CNN architectures. This invention involves the usage of modified CNN with Yolo method for analyzing and marking the Brain malignancy using ADAM (Adaptive Moment estimation) optimization technique. The method extends its scope with modified CNN with modified transfer learning.

BACKGROUND OF THE INVENTION

Artificial Intelligence and its various forms of processing involves wakes up with significant implications for a wide range application to different sectors, including health, social media, and biomedical sciences, robots, aerospace systems, and autonomy. CNNs (Convolutional Neural Networks) can often overcome the human intelligence level based judging the results in various complex problems which includes image processing, audio processing and major statistical data based problem solving consequential long-horizon preparation, thanks to many of its parameters in more counts of millions and a specific lack in the domain-specific expertise. The major aspects of CNN's lie with its data training to identify the features. The CNNs comprises of multiple neurons that could mimic the human brain and it process the data based on the optimisation by learning. In the structure of the

Page 1 of 13

CNNs, the input layer takes part in receiving the data and the major part is handled by the hidden layers. In this processing the weights and biases make the computations. The features of the input data set will be identified and saved to check for the error or the equality. The CNN's can separate or provide isolation and can identify the features that are derived from input images with high accuracy than almost any algorithm, many thanks to this self-optimization characteristic. In addition to these, the quick performing nature of pre-processing the data inputs and the ability to produce acceptable and precise results make them strong in computation. The usage of CNN with different image processing and audio and data processing techniques is known to the world. Many known models of CNNs like Alexnet, Densenet and VGGnet performs better in real world conditions.

Mohammad Irfan Sherif et al discussed about brain tumor segmentation using CNNs. The authors mentioned the task of identifyingthe tumor from the skull via imaging as a challenging task. They mentioned the complexity in imaging process and the detection of the tumor from the MRI. They have a combined algorithmic approach with inception v3 net and Yolo v2 for identification, localization and classification. They have used the database from available intemet sources.

Muhammad Farhan Safdar (2019), discussed about the approach using Machine learning. The researchers discussed in a method to enhance the low volume of the medical imaging dataset, 8 various data augmentation approaches could have been evaluated by using ML with YOLO V3 applied on each dataset individually. There studies resulted that at 180 degrees and 90 degrees the data enhancement techniques could perform better for medical imaging. They have discussed that the crop and scale approach could also help in achieving good results of about 83% accuracy. When compared with the rotation techniques, this accuracy is less. It has been mentioned that some other approaches like synthetic DA by usage of GAB could be explored on medical MRI scans.

Shanchen Pang et al, discussed that to locate diseases exactly from medical images (MRI/CT or Ultrasonic) seems to be of highly challenging and time consuming task. They

Page 2 of 13 commented on the recent advancements in the development of artificial neural networks that promising provides accurate location of the disease from the medical images. They have discussed the case of introducing the Al in cholelithiasis and classifying gallstones on CT images. As the data base from open sources were unavailable, the researchers took the initial step of making a database which was successful with 223846 CT images with gallstone of 1369 patients. From the created data base, a neural network was trained to choose the CT images comprising of good quality for the training process and the Yolo vers 3 is proposed to identify the cholelithiasis and classify gallstones on CT images. This process has undergone a k-folding of 10 times for cross validations for the identification and classification process. The researchers have obtained 92.7% accuracy with this method for the identification of granular kind of gallbladder stones and 80.3% accuracy in identifying muddy kind of gallbladder stones.

SUMMARY OF THE INVENTION

Artificial intelligence being the vast machine learning (ML) and Deep Learning (DL) area is formed by the mimicking of the how a human brain can work. The areas of DL are now seen more commonly Data analysis and Prediction. Along with that Artificial Intelligence (AI) is used in analysis because of its highly precise characteristics like unlabelled data usage, its ability to operate beyond feature engineering, including prediction with higher quality and precision. Wide range of usage of DL includes industries, object tracking, self driving vehicles, facial recognition, and image labelling among many others. The Convolutional Neural Network (CNN) is a deep learning algorithm that has shown to be effective in solving problems such as text processing, image labelling, pose identification, and behaviour recognition. In recent decades, CNN has seen positive findings in the area of medical imaging. CNN is made up of artificial neurons that, like brain neural connections have the ability to self-optimize by learning. It can isolate and identify the features derived from images quite precisely than almost any algorithm contributes to this self-optimizing ability. Furthermore, it needs relatively little pre-processing of every raw data while producing extremely reliable and precise information. The use of CNN in the detection of

Page 3 of 13 object and image processing, like medical imaging, is widespread. VGG19 net, Dense Net, and Resnet 50 are three well-known models of CNN for image categorization that work admirably in real-world situations. Figure 1 shows the input image features if it is MRI or CT data.

It is a known fact that the CNN's can perform better always with a huge dataset. Here the functional data set is limited, but sufficient to carry out the experimentation. The input data base normally under goes the pre-processing operation, where required level of normalization on the input data sets are normally performed. For training the CNN models, it is quite normal in using the data augmentation to transform a minimal data set/group to a much bigger one. There are different types of data augmentation like cutmix, class label type, mosaic data augmentation, self-adversarial training that makes the modified CNN Yolo's more efficient in expanding the data sets. Rather than gathering new data, existing data is getting augmentation, will increase the efficiency of deep learning (DL) models. Augmentation of data is being integrated into many deep learning algorithms. However, this discovery has made three augmentation techniques to produce new training sets. They are scaling, translation, and rotation. The most commonly used one for augmentation is rotation, where an image is rotated in a clockwise manner by an angle from 0 to 360 degrees, which rotates the image's pixel frame and fills the image region where there is no pixilation. A 315-degree rotation (45-degree counter-clockwise) was used in this case. The other augmentation method used is scaling, which seems to be the magnification or lowering of the image's frame scale. A limit of ten percent of the picture is magnified. Image conversion may be performed in either a horizontally (width shift) or vertically (height shift) orientation, sometimes both directions. The original picture was ten percent horizontally transformed and ten percent vertically transformed. Through analyzing the regions of the convolutional layer, enabled in an image and comparing it to the related in the original image's regions, the image's features can be investigated.

A CNN's layers are made up of several 2D arrays known as channels. The output functions of the very initial convolution layer were analyzed after the input picture was added to

Page 4 of 13 various network layers. The input image data after being properly augmented will undergo the image processing using Darknet-19. The modifications are added to the customised deep architecture network called darknet-19. This neural network is originally a 19-layerd network, that has been supplemented with 11 more additional layers for object detection or region of interest detection. Having a 30-layer architecture the modified CNNs with v2 of Yolo found to get struggled with a smaller set of object detection from the image set. These were endorsed to loss of some fine-grained features as when the layers got down sampled the input image. To avoid this, the modified CNN-Yolo v2 can be used to identify mapping concatenating feature maps obtained from the previous layer to capture the low level features in the image. This modified model is optimised for mean square error loss across the prediction bounding and with the ground truth boxes. The experimentation training is carried out on three various types of losses namely, localization, confidence and then classification. Here in this Al based system, the areas are localized and labelled. From the expected and ground truth box, using the location localization loss computes its error. The confidence loss can be used to calculate the objectiveness eroor with the detected object in the 'jth" bounded boz of the grid"i" cell. In the process, the classification loss is utilized to calculate the probability seen along each class of grid cell "i". The experimental outcomes helps in classification of the Brain images, whether it is normal or malignant. The brain images with malignancy is localized using this modified method. The ground truth labelling for the malignancy in the brain image data sample is shown in Figure 2. The modified CNN with modified transfer learning with best suited optimization method could also classify the datasets. The modifications in the selected CNN's at their fully connected network layers and the optimization techniques will provide the necessary classifications. The training iterations taken more, will enhance the accuracy of the model and could provide better results. Figure 3 shows the proposed model where the key representation is the Modified CNN. The proposed model uses ADAM (Adaptive Moment estimation) optimization methods whose importance is described in the section below. From this part of CNN's modification itself, the selection can be with classification and then malignancy marking by modified CNN with Yolo and the next option is the model can use modified transfer learning methods for classification.

Page 5 of 13

DETAILED DESCRIPTION OF THE INVENTION

Embodiments, of the present disclosure, will now be described with reference to the accompanying drawings.

Embodiments are provided so as to thoroughly and fully convey the scope of the present disclosure to the person skilled in the art. Numerous details are set forth, relating to specific components, and methods, to provide a complete understanding of embodiments of the present disclosure. It will be apparent to the person skilled in the art that the details provided in the embodiments should not be construed to limit the scope of the present disclosure. In some embodiments, well-known processes, well-known apparatus structures, and well-known techniques are not described in detail.

The terminology used, in the present disclosure, is only for the purpose of explaining a particular embodiment and such terminology shall not be considered to limit the scope of the present disclosure. As used in the present disclosure, the forms "a," "an," and "the" may be intended to include the plural forms as well, unless the context clearly suggests otherwise. The terms "comprises," "comprising," "including," and "having," are open ended transitional phrases and therefore specify the presence of stated features, integers, steps, operations, elements, modules, units and/or components, but do not forbid the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof The particular order of steps disclosed in the method and process of the present disclosure is not to be construed as necessarily requiring their performance as described or illustrated. It is also to be understood that additional or alternative steps maybe employed.

The proposed modified Convolutional Neural Network based learning algorithmic approach is enhanced in Brain MRI/CT malignancy image identification or diagnosis. Convolutional Neural Networks are neural networks that mimic the human brain. The

Page 6 of 13 modified CNN with Yolo-v2 is resembling like a fully convolutional neural network (FCNN). This network when fed with the input, will forward the image which is (n * n) into the FCNN and will provide an output of (m*m) prediction. Using this modified CNN Yolo architecture, its splits the input data image into (m*m) grid. Further for each generation of grid, two bounding boxes along with probability of those classes are made. It is noted that the bounding box is slightly larger than the grid generated. The proposed modified CNN-Yolo trains on the Brain MRI/CT images with and without malignancy. It directly optimizes detection performances. This method holds advantages over the conventional methods used for object detection.

The main factor of modified CNN-Yolo is it is quick in computation as it is framed as a regression problem and hence does not require a complex route. The CNN's process need to just run the neural network with a new test image to check for the prediction in the diagnosis. The base model of the network is capable of running a min of 45 frames per second without a batch processing in a workstation with RTX 3000. The next or the second noted thing about the modified CNN-Yolo is that, it reasons or could search widely about the input brain images while making a decision for the predictions. Not like the region proposal based and sliding window techniques, this modified CNN-Yolo checks the full image in its training process as well as in test time, so as it could discretely encode some of the contextual information about various classes along with its appearance. Eg: The fast R CNN know as one of the best object detection method, wrongly takes the background patches in a given input image due to the reason that it cannot identify the larger context. Here, this CNN-Yolo-v2 could help in avoiding such kind of errors when in comparison to the Fast R-CNN. The next main nature of modified CNN-Yolo is that it could learn generalizable depictions of input data/objects. Once if the network is trained on natural or normal images and then testing is performed eg: on an artwork, then this proposed modified CNN-Yolo will outperform some other high rated object detection methods like R-CNN and DPM by a huge margin. Hence this proposed Al based system is highly generalized. The proposed neural network thus uses these features from the whole input image to finalize a prediction with each bounding box. This can also induce a prediction with all

Page 7 of 13 bounding boxes that is spread over all classes of an input brain image concurrently, which means this network could reason widely on a complete image and the specified objects in the input image. This modified CNN enables to have a fully tracked training within and could offer quick computation while managing the precision.

The proposed modified CNN model involved in categorizing the MRI/CT includes dataset collection, data pre-processing, dataset classification, training models followed by evaluation, and finally, the model is analyzed. Initially, the dataset is required, and therefore, it is collected and arranged to provide the train and validate the proposed system model. To acquire the unity in data, the collected data are altered, resized, and regularized. Under similar environmental conditions, the dataset of all models is trained and validated.

The proposed system is performing the process by dividing the input data image - ie Brain MRI/CT images into D*D grid from. When the center of an object meets the criteria and gets into the grid cell, then that particular grid cell will be responsible in detecting that image/object. Here, each grid cell will be able to predict the bounding box, B and their confidence score. In depth, these confidence scores will reflect on how confident is this proposed model and the accuracy in prediction. The confidence parameter is checked and set as by Pr(Object)*IoU, i.e, Intersection over Union. If by chance, there exists no objects in that particular cell, then the confidence score be indicated as zero.

The details in the process is that, five predictions are set for each bounding box : x, y, w, h and confidence parameter. The coordinates x and y indicates the center of the box with respect to the boundary of the grid cell. The width w and height h could be predicted in equivalence to the ll input image. At the final stage the confidence prediction will present the Intersection over Union among the ground truth box and prediction box. In this function, the ability of each grid cell to predict the conditional probability class, C is noted as impressive. All these possibilities/probabilities can be conditioned on based on the grid cell that contains the required object or region of interest. Regardless of the number of boundary boxes, there shall only be a single set of class probabilities for a grid cell. The

Page 8 of 13 proposed method uses ADAM (Adaptive Moment estimation) optimization method, which is a replacement optimization algorithm to replace stochastic gradient descent to train the deep learning models. The advantage of using this ADAM is that, it could combine the very best characteristics of the AdaGrad and RMSPro algorithms which can provide an optimization algorithm that could grip sparse gradients on noise affected problems. The next concept of choosing ADAM is due to its relatively easier method to configure in which most of the parameters get configured by default and could perform well.

Thus the proposed modified CNN-Yolo sets for Dataset 1 for normal subject and Dataset 2 for malignancy affected is acquired for the process to assist the technicians in the neuro imaging field to accelerated the diagnosis process. The proposed CNN model is trained and validated using both the Datasets. With the trained data sets, and evaluation for validations, new and random images where checked and found to be identified and marked as malignant by the Al based system.

Page 9 of 13

TITLE OF THE INVENTION: "Al SYSTEM FOR BRAIN MRI/CT MALIGNANCY IDENTIFICATION AND CLASSIFICATION USING MODIFIED CNN AND ADAM OPTIMIZATION." CLAIMS,

We Claims,

[CLAIM 1] The proposed modified CNN based learning algorithmic approach is enhanced in brain MRI/CT image classification. i. From claim 1, the CNN network model is used in object recognition and image processing, like digital diagnostics. This invention is also intended in the usage of the modified CNN-Yolo-v2 Learning algorithm concept, which uses convolutional layers and fully connected layers for input data processing. The major processing is performed by the arrangements of fully connected neural networks. The CNN-Yolo-v2's part is mostly used while labelling the Data for training. ii. From claim 1, this model accurately identifies the Subject's MRI/CT with the high precision with very less effort and high accuracy. iii. From claim 1, AGDM optimization is used in the proposed system is to achieve the consistent gradient by enabling a high dimensional velocity with the simultaneous reduction in problems.

[CLAIM 2] A Al based modified CNN Yolo v2 -system model is presented in this innovation to differentiate the affected subjects through MRI/CT scans. i. From claim 2, the faster Yolo-CNN comprises of artificial neurons which mimic the human brain's neural connections and could perform self-optimization by learning. ii. From claim 2, the proposed Yolo-CNN model involved in classifying

Page 1 of 2 the brain MRI/CT involves collection of the training input data, pre processing of the data for system specification, classification or diagnosis of the data, data training, it's evaluation, and analysis. iii. From claim 2, the proposed modified CNN-Yolo v2-model is the quickest and appropriate method in detecting Brain malignancy problems and further could support the diagnosis and treatment with a higher accuracy rate.

[CLAIM 3] The modified CNN Yolo v2-learning will train the data set and will mark the data with the malignancy and further complete the training process. i. vii. From claim 3, the Brain MRI/CT images of patients are collected from the open source libraries like Kaggle/Github repository etc are used for training and validation purpose for the proposed model. ii. From claim 3, the large dataset with normal subjects and affected (malignancy) is utilized in the training process with majority chosen for training the Al system while a 20% or 30% data are acquired for the process ofvalidation. iii. From claim 3, the result shows that the proposed model achieves a proper marking of the malignancy in the Brain images using the modified CNN Yolo v2 method.

[CLAIM 4] The modified CNN with modified transfer-learning can also train the data set and validate the data using the given multiple categorized data sets. i. From claim 4, the MRI/CT data set collected from the open source libraries can be used for training and validation, in a 70:30 percentage mode for the next method of CNN proposed model. ii. From claim 4, Modifications in the fully connected neural network layers can yield proper classification. Dated this 1 9 th August 2021.

Page 2 of 2

AI SYSTEM FOR BRAIN MRI/CT MALIGNANCY IDENTIFICATION AND CLASSIFICATION USING MODIFIED CNN AND ADAM OPTIMIZATION FIGURES

FIGURE 1: MRI AND CT IMAGING DETAILS 2021106727

FIGURE 2: SAMPLE OF AI GROUND TRUTH LABELLING OF DATA FOR BRAIN MALIGNANCY

FIGURE 3: THE PROPOSED SYSTEM MODEL