Robust and Flexible Deep Learning Models with Incomplete Medical Image Datasets

Abstract

Deep learning has revolutionized the field of medical image analysis in the past decade. The success of deep learning can be attributed to the availability of fully annotated medical image datasets. However, the real-world medical image datasets are often heterogeneous and incomplete, such as images with missing modality and partial annotations. Such incomplete datasets significantly hinder the development and deployment of deep learning models for practical use. In this dissertation, several innovative techniques are developed to build robust and flexible deep learning models by exploiting incomplete medical imaging datasets. (1) I propose a generic technique to improve the model robustness to handle incomplete modalities for multi-modality MRI. (2) I explore the feasibility of using synthetic images to diminish the demand of a certain modality for multi-modality imaging. (3) I develop a deep-learning-based landmark localization model for deep brain stimulation with limited human annotation. A comparative study is performed to further evaluate the localization performances of the automatic approaches and the inter and intra-rater variability. (4) To transfer knowledge from one modality to another, I investigate the techniques for unsupervised cross-modality domain adaptation, especially image-level domain alignment. To address the intra-domain variability, I introduce several novel unpaired image translation approaches to improve the style diversity of the translated images. (5) To unleash the synergistic potential of partially labeled datasets, I develop a novel partial label segmentation technique by combining different supervision signals efficiently and effectively. (6) I explore efficient annotation strategies for 3D medical image segmentation by conducting a benchmark study for cold-start active learning approaches.