NOVEL EFFICIENT DEEP LEARNING ARCHITECTURES FOR MEDICAL IMAGE ANALYSIS

Abstract

Medical image analysis is now a significant field of research in the health care domain, because the requirement of more accurate, automated and efficient computational diagnosis systems. It is recently observed that the deep learning methods achieve remarkable success on medical image classification, segmentation, and disease detection, however, existing deep learning models are extremely complex and have high memory and runtime demand that can't be easily used in clinic practice. Beside, the anatomical consistency on medical image segmentation, particularly on 3D medical image applications, is still a challenge. In this thesis, the study develops new efficient deep learning architectures for medical image analysis, based on the investigation of the issues of computational efficiency, representation learning, and anatomical consistent segmentation. Two approaches are introduced: first approach propose a sparse attention based contrastive learning framework on medical image representation learning. Sparse attention helps reducing computational overhead and enhances feature extraction and learning from medical image. Contrastive learning is used to learn discriminative representation. Second approach proposed a topology constrained quantized nnU-Net framework for efficient and anatomically consistent 3D dental image segmentation. By adding topological constraints, anatomy consistency on segmentation is preserved while quantization on the model makes it computationally efficient and reduces model complexity and usage of hardware memory to be easily deployed in real application environments with less loss in segmentation performance. Experimental results show that the proposed models have comparable accuracy but much better computational efficiency and robust performance in terms of learning ability and structural accuracy in dental images segmentation applications. Therefore, these methods have important potential applications in automatic diagnosis system, clinical intelligent support systems, auxiliary system for radiologist, efficient healthcare systems, and so on. The potential future directions are multimodal medical imaging applications, federated learning, explainable ai, and lightweight deployment-oriented models for real time applications.