MACHINE LEARNING FOR MULTI-LABEL SEGMENTATION OF MEDICAL IMAGES

Abstract

The objective of medical image segmentation is to partition the image into several meaningful segments to represent different tissues, organs, pathologies or other biologically relevant structures. Although segmentation is significant in diagnosis and treatment planning, manual labeling is tedious and time-consuming. The development of approaches for automated segmentation faces challenges such as class imbalance and inter-class interference. In this thesis, several efficient pipelines that adopt machine learning techniques for medical image segmentation, especially the multi-label segmentation scenarios, are proposed. Instead of relying on heavy model ensembles or complicated post-processing steps, the proposed approaches exploit the power of discriminative feature representation, coarse-to-fine segmentation and the characteristics of a specific disease to improve performance. They achieve competitive performances relative to state-of-the-art methods without sacrificing practical value for clinical use.