From Data-driven to Data-centric Medical Image Segmentation

Abstract

In this dissertation, I develop deep learning methods for robust medical image segmentation, transitioning from data-driven to data-centric approaches. This shift highlights the critical importance of data quality in improving segmentation efficacy. While data-driven strategies focus on improving neural network (NN) performance with existing datasets, data- centric methods emphasize the essential role of data quality and diversity in improving segmentation performance. The data-driven aspect involves proposing state-of-the-art NNs for robust segmentation in various medical applications. The data-centric aspect has two key components: addressing domain shifts and incorporating domain knowledge from experts. To tackle domain shifts, I improve data quality, consistency, and diversity via unsupervised domain adaptation and test-time adaptation. Additionally, incorporating domain knowledge from human experts strengthens model robustness and adaptability, improving accurate and reliable segmentation across different challenging cases. The proposed data-driven and data- centric medical image segmentation methods have demonstrated superior performance, producing robust outcomes across a variety of tasks, imaging modalities, and populations.