Deep learning-based computer-aided diagnosis for brain and retinal diseases

Deep neural network has achieved excellent performance for many recognition tasks. Despite its recent wide application on medical imaging tasks, the requirement of large amount of manually labeled samples limits its performance on medical image recognition tasks. Comparing with natural images, medical image is difficult and expensive to acquire and requires specialized training for its labeling. However, the data samples for a specific clinical task shares much less heterogeneity comparing with most image recognition tasks. Exploring advanced network architecture and incorporate it with a-priori, domain-specific knowledge has a great potential to deliver superior recognition performance and better computer-aided diagnosis system. In this thesis, we presented the development of four novel deep learning based frameworks regarding four medical image recognition tasks, early diagnosis of Alzheimer's disease, differential diagnosis of multiclass dementia, OCT retinal fluid segmentation and OCT retinal layer segmentation. Comprehensive experiments proved that the proposed frameworks out-performed state-of-the-art methods in each individual task.