Towards Resilient Models: A Deep Learning Odyssey through Mammographic Images
Loading...
Authors
SARKER, SUSHMITA
Issue Date
2023
Type
Thesis
Language
Keywords
Breast Mass Classification , Breast Mass Segmentation , Deep Learning , Mammogram , Multi-view
Alternative Title
Abstract
Recent advancements in deep learning have revolutionized the landscape of breast cancer diagnosis and analysis. While these strides have been remarkable, challenges persist in developing robust and accurate models tailored specifically for medical imaging tasks. This thesis explores the complexities of this domain by addressing two pivotal aspects: mammographic image classification and segmentation. In contrast to conventional single-view analyses, this approach recognizes and harnesses the intrinsic correlations among the four views in a mammography exam. An innovative multi-view network based on transformers is introduced, specifically crafted to elevate mammographic image classification. The contribution includes a distinctive shifted window-based dynamic attention block, leveraging transformers to seamlessly integrate multi-view information. This facilitates efficient information transfer between views at the spatial feature map level, resulting in a substantial improvement in tumor detection. To further investigate the impact of Convolutional Neural Network (CNN) architectures for medical image segmentation, two novel CNN architectures, ConnectedUNets+ and ConnectedUNets++ are proposed, each tailored for single-view segmentation. Additionally, for multi-view segmentation, an enhanced Cross-View Attention Network block is introduced. Collectively, this thesis contributes to the refinement and advancement of deep learning models for breast cancer diagnosis and analysis, with a specific focus on multi-view classification and segmentation tasks. The proposed methodologies exhibit promising results and pave the way for future developments in the field of medical image analysis.