Treffer: Weakly supervised deep learning for large-scale medical image analysis

Ph.D. ; The computer-aid diagnosis (CAD) plays an essential role in clinical practice and has been a long-standing topic in medical image analysis. To develop a robust and effective CAD system, a great number of fine-grained annotations that indicate local diagnostic information are indispensable. However, the acquisition of such annotations is always expensive, time-consuming, and sometimes even impractical. Compared to massive fine-grained annotations, coarse labels (e.g., image-level labels) or a handful of fine-grained annotations are much easier to obtain from clinical records or experts. How to maximize the utilization efficiency of available medical data to improve diagnostic performance is the general scope of weakly supervised learning (WSL) methods, which can substantially reduce experts' annotation efforts. ; This thesis presents a series of weakly supervised deep learning approaches to achieve effective learning from coarsely-annotated or limited labeled training data, aiming at building generalized and intelligent CAD systems. In the first two parts, we address the classification problem of two types of large-scale images (gigapixel histopathology images, and high-dimensional optical coherence tomography (OCT) scans) by learning from coarse annotations under inexact supervision. In the remaining parts, we study semi-supervised deep learning from limited labeled data and plenty of unlabeled data under incomplete supervision for tackling issues of missing labels or insufficient fine-grained labels. ; In the first part, we focus on the problem of multi-class lung cancer whole slide image (WSI) classification. The automatic classification of WSI remains challenging owing to the formidable image size, the diversity of biological structure, and the sparse distribution of tumors. To overcome these challenges, we propose a WSL framework for fast and effective classification. We leverage a patch-based fully convolutional neural network (CNN) for retrieving discriminative blocks and providing representative ...