DEEP LEARNING-DRIVEN APPROACHES TO MEDICAL IMAGE COMPUTING AND SECURE PROCESSING

Abstract

This thesis presents a comprehensive exploration of the integration of machine learning techniques in medical imaging, a field witnessing rapid transformation due to technological advancements. The primary focus lies in enhancing the precision and effectiveness of medical image processing, which bridges the gap between computational algorithms and practical medical applications. The research encompasses several key areas: advanced medical image segmentation, data security and privacy in medical imaging, innovative data representations, and the development of resource-efficient learning models for medical diagnostics. A significant contribution of this work is the refinement of the U-Net architecture for medical image segmentation. By innovatively implementing efficient neural connections and strategic skip paths, this research significantly reduces the semantic gap in concatenated feature maps, resulting in enhanced accuracy and precision. This methodology is versatile, applicable across various medical imaging modalities, and particularly effective in skin lesion segmentation. In addressing the crucial issue of data security in medical imaging, this thesis introduces a dual encryption method that combines the strengths of blowfish encryption and certificateless signcryption. This novel approach offers a secure, computationally efficient solution to protect the integrity and confidentiality of sensitive medical data during transmission, surpassing traditional security measures.Another pioneering aspect of the research is the transition from traditional 2D imaging techniques to 3D point cloud representations. This approach provides a more comprehensive and detailed visualization of medical conditions, utilizing saturated point clouds and hypersurface constructions. Additionally, the application of the wavelet transform for multi-scale image processing marks a significant advancement in enhancing diagnostic capabilities. Complementing these innovations, the thesis also presents a resource-efficient machine learning model tailored for medical image analysis. This model achieves a delicate balance between diagnostic accuracy and computational efficiency. By integrating advanced image processing techniques such as anisotropic diffusion and tumor region augmentations, the model ensures realistic variations in training data, thereby improving its diagnostic precision. Overall, this research represents a rigorous and multifaceted approach to applying machine learning algorithms in medical image processing. It contributes to both the fields of computer science and medical science, offering novel methodologies and insights that have profound implications for diagnostics and treatment planning.