Treffer: Maximizing the data utilization efficiency in medical imaging diagnosis: from full supervision to weak supervision

Ph.D. ; The annotated data provided by experienced experts is at the core of AI-powered automatic medical imaging diagnosis systems. It is expensive and hard to obtain on a large scale, particularly in the field of medical imaging, where only domain experts, i.e., radiologists, can provide reliable annotations. In this thesis, the effective usage of the limited data is called data utilization efficiency}. To improve the diagnostic performance of medical imaging diagnosis systems, maximizing the data utilization efficiency is very important. In this thesis, we study with this main goal from full supervision to weak supervision, to address some common challenging problems in the medical image diagnosis field. ; The computer-aided diagnosis (CAD) has been a long-standing topic in medical imaging computing, including automatic disease diagnosis and abnormal region segmentation from various medical images, e.g., multi-modality MRI, CT scans, dermoscopy images, retinal fundus images, etc. These computer-aided diagnosis techniques show the interpretation of medical images and quantitative measurements, which can assist doctors in determining a more accurate diagnosis or treatment planning. With the availability of the massive amount of data and annotations, deep learning has become a de facto standard approach in a variety of medical image diagnosis applications. ; In the first part, we tackle typical and challenging problems in CAD under full supervision. Toward a more precise diagnosis, we propose several methods to better utilize the data to extract discriminative features for recognition. First, we propose a 3D multi-scale fully convolutional network (MsFCN) with random modality voxel dropout (RMVD) for automatic intervertebral disc (IVD) localization and segmentation. Our method incorporates multiple scales of IVD in the network and alleviates the co-adaptation issue in multi-modality images, thus contributes to better performance. Second, 3D FCNs would suffer from large computational inefficiency, especially ...