Learning with imperfect datasets in medical image segmentation

edical image segmentation aims to partition medical images into distinct physiological regions, such as organs and lesions, which is crucial for disease diagnosis and treatment planning.The advent of deep neural networks has significantly advanced this field.However, the performance in real-world scenarios remains unsatisfactory due to imperfect data and annotations during both training and deployment.First, scaling up training data and annotations is challenging.This is because obtaining medical images is difficult due to privacy concerns.Furthermore, annotating medical images requires substantial expertise, making the process costly and difficult to carry out.Second, the quality of medical images cannot always be guaranteed in real-world scenarios, leading to significant performance drops in outlier cases.Third, real-world medical applications are safety-critical and demand extremely accurate predictions, a requirement that most existing models fail to meet adequately.These challenges hinder the practical deployment of medical image segmentation.Consequently, both academic and industrial communities are striving to develop highly accurate medical image segmentation algorithms that can perform well despite imperfect data.To this end, this thesis proposes deep learning methods to develop a well-performed medical image segmentation model that can be effectively trained with limited and low-quality data/annotations.Specifically, a comprehensive suite is proposed from three directions: (1) applying image registration to generate realistic and diverse training samples and adopting barely-supervised learning paradigms to enable learning with insufficient annotated data; (2) devising region-aware fusion module to address missing modality problem; (3) incorporating automatic and interactive medical image segmentation into a single model and one training session to achieve sufficient segmentation performance for practical use.Extensive experiments on several medical image segmentation tasks, such as brain tumor segmentation, brain structure segmentation and abdominal organ segmentation, demonstrate the effectiveness and efficiency of the proposed techniques.ix