Thin-Thick Adapter: Segmenting Thin Scans Using Thick Annotations

Medical imaging segmentation has been a prominent focus in the field of medical imaging analysis. Recent advances in radiological and storage technologies have led to an increased utilization of thin slice computed tomography (CT) acquisitions in clinical practice. These thin slices offer several advantages, including enhanced spatial resolution and sharper diagnostic information for clinicians. However, segmenting thin slices presents significant challenges. Annotations on thick is hard to adapt to the thin slices since there is a domain gap between thick and thin slices. Furthermore, there is no existing dataset which contains pixel-level thin annotations, and manually annotating thin slices is considerably more resource-intensive and time-consuming compared to annotating thick slices, making it impractical to obtain a sufficient quantity of high-quality thin annotations for training robust models in a supervised fashion. In response to these challenges, this paper introduces three key contributions. Firstly, we propose a research topic and setting focused on segmenting thin slice data exclusively, leveraging existing annotations from thick slices. Secondly, we present a newly created dataset called CQ500-Thin, which is a Non-Contrast CT scans featuring Intracranial Hemorrhage (ICH), including a subset of pixel-level thin annotations for evaluation purposes. This dataset serves as a benchmark for our proposed topic and methodology. Lastly, we introduce a robust pipeline named the Thin-Thick Adapter, which utilizes a simple-but-effective data alignment technique and a 3D-CPS for unsupervised domain adaptation. It is designed to address the thin slice segmentation problem and establish a foundational baseline for this emerging research area.