Application of deep learning to optimize computer-aided-detection and diagnosis of medical images

Abstract

The field of medical imaging informatics has experienced significant advancements with the integration of artificial intelligence (AI), especially in tasks like detecting abnormalities in retinal fundus images. This dissertation focuses on four interrelated research contributions that address crucial aspects of AI in medical imaging, offering a comprehensive overview of various innovative approaches and methodologies. The first contribution involves developing a two-stage deep learning model. This model significantly improves the accuracy of identifying high-quality retinal fundus images by eliminating those with severe artifacts. It highlights the critical role of an optimal training dataset in enhancing the performance of deep learning models. The second contribution presents an innovative algorithm for synthetic data generation. This algorithm enhances the effectiveness of deep learning models in medical image analysis by augmenting datasets with synthesized annotated diseased regions onto disease-free images, leading to notable improvements in disease classification accuracy. The third contribution is centered around a novel joint deep-learning model for medical image segmentation and classification. Combining a U-net architecture with an image classification model it demonstrates substantial accuracy improvements as the training dataset size increases. Lastly, a comparative analysis is conducted between radionics-based and deep transfer learning-based Computer-Aided Detection (CAD) schemes for classifying breast lesions in digital mammograms. The findings reveal the superiority of deep transfer learning methods in achieving higher classification accuracy. Collectively, these contributions offer valuable insights and practical methodologies for enhancing the efficiency and diagnostic accuracy of AI applications in medical imaging, marking a significant step forward in this rapidly evolving field.