Advancing medical image segmentation via exploiting limited annotations

In medical image segmentation tasks, most supervised learning algorithms are driven by sufficient annotations, which are usually expensive and labor-intensive to obtain. Therefore, recent deep- and machine-learning advances have sparked a trend toward more data-efficient techniques. Aiming to mitigate the data-hungry issue and improve the model performance, we propose to advance medical image segmentation by exploiting limited annotation. This chapter introduces two of our proposed data-efficient frameworks, respectively, following the idea of partially supervised learning and synthetic data augmentation with generative models. Although implemented with different schemes, they share the same aim to make the most use of available annotated data. The specific techniques in each task can be different, including prior knowledge extraction, partial-annotation integration, and domain translation, among others. We tested our frameworks on various medical image segmentation datasets, and the results show that our proposed frameworks generally improve the model performance. We hope our works can be promising solutions to the data-hungry nature of medical image segmentation tasks.