Transforming Medical Image Segmentation with Enhanced U-Net Architectures and Adaptive Transfer Learning

Abstract

Medical imaging has revolutionized healthcare by enabling accurate diagnosis, treatment planning, and monitoring of various diseases. Various modalities, such as computed tomography (CT), magnetic resonance imaging (MRI), and ultrasound, visualize diverse anatomical structures and pathological conditions. However, challenges arise in medical image segmentation due to increasing complexity, variability, noise, artifacts, and scarcity of annotated data. The advent of AI, particularly deep learning with Convolutional Neural Networks (CNNs), has facilitated significant advancements in medical image segmentation. U-Net, a prominent CNN architecture, provides accurate segmentation results with relatively low training samples due to its encoder-decoder structure with skip connections. In addition, transfer learning further mitigates limitations imposed by scarce labeled data. In this thesis, we develop innovative custom U-Net models with advanced building blocks and transfer learning strategies, such as AM-UNet for human brain claustrum segmentation from MRI scans, TLU-Net for organ and tumor segmentation from CT scans, and OCU-Net for oral cancer tissue segmentation from whole slide images (WSI) stained with Hematoxylin and Eosin (H&E). Furthermore, we introduce the "U-Framework", a comprehensive guide in designing and optimizing U-Net models. This framework encompasses key decisions related to architecture, transfer learning, module selection and fine-tuning, and evaluation strategies. Finally, by comparing our models with state-of-the-art approaches on benchmark datasets, we demonstrate their significant potential to contribute to medical image segmentation.