TOWARDS AUTOMATED AND ANNOTATION-EFFICIENT MEDICAL IMAGE ANALYSIS

Abstract

With the global expansion of medical imaging, the volume of acquired medical image data is increasing at a pace much faster than what human experts can interpret. This thesis is devoted to developing automated and annotation-efficient medical image analysis algorithms to tackle above real-world problem. The thesis consists of two parts, in the first part, I worked with a research team from National University Hospital to develop a deep learning system for stenosis diagnosis at lumbar spine MRI. In the second part, I conduct research on annotation-efficient learning algorithms. I investigate on the potentials of auxiliary labeled, noisy labeled, partially labeled, and unlabeled datasets for annotation-efficient learning. All the proposed algorithms and new learning paradigms help alleviate the annotation burden significantly and facilitate wider adoption of learning solutions in healthcare. Finally, I integrate all the proposed algorithms into GEMINI platform for domain experts to use them at ease.