ANALYSIS OF MEDICAL IMAGES USING AI TECHNIQUES

Abstract

This thesis provides a comprehensive examination of state-of-the-art techniques for detecting brain tumors using Magnetic Resonance Imaging (MRI) images. The research thoroughly examines various segmentation and classification methods and proposes innovative approaches to address the challenges associated with processing medical images. Using various deep learning models, such as 3D Convolutional Neural Networks (CNNs), ResNet, U-Net, and ViTs, the study aims to enhance brain tumor detection efficiency and accuracy. Chapter 2 delves into the most sophisticated segmentation and classification methods used in brain MRI analysis. It examines the strengths and weaknesses of various techniques, including Sine-Cosine Fitness Grey Wolf Optimization, cascade CNNs, and 3D CNNs. The chapter highlights the importance of finding a balance between computational complexity and the need for extensive training data to ensure the generalizability of results across different datasets. In Chapter 3, the noise and distortions deteriorate the effectiveness of MRI scans and also reduce their image quality. Transmission system interference and poor imaging are the main causes of this distortion, typically corrupted by Rician noise (during acquisition). This work focuses on eliminating noise from MRI during its preprocessing phase using a Deep Neural Network (DNN). Image features are extracted from the noisy images by employing a series of convolutions. The proposed model shows an arrangement of encoder-decoder to preserve important image features while rejecting the ones that are not important. Various denoising techniques like average filter, median filter, wavelet smoothing filters, bilateral filters, Gaussian filter, and non-local mean filter are compared to the proposed technique using evaluation parameters like Root Mean Square Error (RMSE), Peak Signal-To-Noise Ratio (PSNR), and Mean Square Error (MSE). The results show that the proposed technique was successful in preserving essential image features. In addition to the proposed DNN, a comprehensive pipeline for brain tumor identification is presented, commencing with the denoising method known as Modified Triplet Cross Fusion Learning (MTCL-MRI). This novel method dramatically lowers noise while maintaining crucial structural details, enhancing the vi clarity of magnetic resonance imaging. On the BraTS 2020 dataset, segmentation with U-Net and classification using ResNet-based techniques produced a classification accuracy of 99.56%, surpassing state-of-the-art methods and indicating the possibility of improved diagnostic accuracy. In Chapter 4, transfer learning is applied to the classification of brain MRI images using six pre-trained CNN models (VGG16, VGG19, InceptionV3, ResNet50, ResNet101, EfficientNetB1). After these models were fine-tuned and evaluated, VGG16 and ResNet101 were found to be the best models. They achieved the highest classification accuracy, demonstrating the effectiveness of transfer learning in raising training accuracy and efficiency. In Chapter 5, an innovative approach to overcome the drawbacks of conventional downsampling techniques is presented. It combines triplanar isolated spatial pyramidal pooling with Atrous convolution. This method preserves important local and global contextual information, lowering computational costs while increasing detection accuracy. Superior performance metrics, including Dice scores and mean intersection over union (mIoU), were demonstrated by the method, underscoring its potential for precise and early tumor detection. The HyperSwinNet framework, which combines self-supervised pre-training with a transformer-based architecture, is introduced in Chapter 6. With a Dice score of 90.4% on average on the BraTS dataset, this framework effectively traverses the large segmentation network search space. Its clinical relevance and potential to improve diagnostic and therapeutic outcomes are highlighted by its robustness and precision, which have been validated through a variety of MRI tasks. In order to address the problems of limited data and lengthy computation times, Chapter 7 uses a Vision Transformer (ViT) with a standalone GAN for preprocessing, normalization, and pixel segmentation. The model's effectiveness in extracting global and local features, improving image similarity, and streamlining the training process is demonstrated by its high accuracy and sensitivity values. In Chapter 8, a hybrid intelligent system that blends CNNs and recurrent neural networks (RNNs)—more specifically, Long Short-Term Memory networks—is used to investigate a multi-mark classification method. By adding multiple labels to different MRI scan sections, this method increases diagnostic correctness and vii significantly improves performance metrics like mean square error (MSE) and probability of occurrence (POC). This thesis's research demonstrates important developments in the area of MRI-based brain tumor detection. The work improves patient outcomes and diagnostic accuracy by tackling important issues like precise classification, accurate segmentation, and noise reduction. The potential of these techniques to revolutionize medical imaging and progress the field of brain tumor detection is highlighted by their inventive methodologies and thorough evaluations. Subsequent research endeavors will center on optimizing these techniques, broadening their scope, and tackling residual obstacles to ultimately augment the effectiveness of brain tumor identification through cutting- edge deep learning technologies. This thesis comprehends several promising future scopes for enhancing the rate of brain tumor detection using MRI scans. By putting together multi-modal imaging, a more thorough diagnostic approach can be provided. Accurate and timely decision- making can be achieved by implementing these models in real time. By examining tumor characteristics, a customized treatment plan can be developed. Moreover, image structure details can be protected and enhanced by advancements in image preprocessing, such as normalization and denoising by using various advanced techniques. In medical image processing, the limitation in training data is one major drawback that can be overcome by generating synthetic data using advanced AI models. The knowledge of tumor progression can help in improving the effectiveness of treatment when talking about brain tumors. This can be achieved using longitudinal and hybrid models. After all, a more effective, accurate, and clinically feasible treatment is the goal.