Abstract
Histopathological image analysis is challenging but essential for clinical disease detection, diagnosis, and prognosis prediction. Machine learning-powered computer-aided diagnosis (CAD) systems could significantly improve pathologists’ performance and provide evidence to support their decision-making process. However, the conventional single-task-based machine learning algorithms achieved poor generalizability for histopathology image analysis, particularly when only limited data is available. In this dissertation, I build a suite of deep multi-task learning (MTL) approaches to enhance the generalizability of the models for the most challenging tasks in histopathology image processing. MTL uses multiple domain information-enriched tasks as inductive biases to improve the accuracy and interpretability of machine learning models. It enables machine learning models to learn shared representation from multiple tasks to exploit the commonalities and differences between them, which could greatly reduce models’ dependency on large datasets. First, a bending loss-regularized MTL network is proposed to accurately extract overlapped nuclei from histopathology images. Second, I propose a topology-aware network that simultaneously performs instance segmentation and gland topology estimation, which segments severely deformed and densely clustered glands accurately. Third, a novel generative network is proposed to synthesize histopathology images; and it generates realistic images from the layout to the target styles that align with different organs and cancer. Lastly, I explore MTL approaches for other image modalities in real-world problems. I present convincing visuals in many histopathology image applications, including nuclei segmentation, gland segmentation, histopathology image generation, and realistic manipulation of textures.