Deep Learning Based Segmentation and Classification of Medical Images

Abstract

Deep Learning (DL) methods are revolutionizing Medical Image Analysis (MIA), particularly in segmentation and classification tasks. Traditionally reliant on radiologists, medical image interpretation can be error-prone and time-consuming. DL excels in handling complex datasets, positioning it as a crucial tool for enhancing MIA by compensating for the scarcity of radiological expertise and advancing adaptive, efficient, and patient-centered diagnostics. newlineMedical imaging faces significant hurdles due to the specialized nature of data collection and inherent imbalances in dataset distribution, which often leads to biased model outcomes. Addressing these challenges, our research focuses on innovative data augmentation using Generative Adversarial Networks (GANs) and domain-aware Transfer Learning (TL). GANs help augment dataset diversity without the ethical issues of direct data collection, while domain-aware TL customizes the learning process to the specifics of medical images, enhancing model relevance and performance. newlineOur work also introduces a novel U-Net and Vision Transformer (ViT) hybrid model for segmentation, combining U-Net s effective local feature extraction with ViT s global contextual understanding. This hybrid model aims to improve segmentation accuracy, especially in complex scenarios traditional methods struggle with due to their limited scope of context integration. newlineThe research extends beyond conventional methods by integrating modifications such as a novel set of layers in the transfer head network, an adjusted loss function incorporating class weights, and the use of cyclically varied learning rates. These innovations contribute to a more robust, efficient, and effective diagnostic process, suited to the dynamic needs of medical imaging. newlineOverall, by addressing the gaps in current methodologies and integrating advanced DL strategies, this thesis proposes new models and techniques that significantly enhance the accuracy and efficiency of medical diagnostics.