Learning From Target-Level Incomplete Annotation: A Novel Perspective for Weakly-Supervised Multi-Lesion Segmentation

Accurately segmenting various clinically significant lesion areas from whole-body computed tomography (CT) scans is crucial for automated diagnosis and treatment planning. Training an automatic segmentation model effectively is desirable, but it heavily relies on a large scale of pixel-wise labeled data, which is laborious, time-consuming, and expensive to obtain. Existing weakly-supervised segmentation approaches often struggle with regions nearby the lesion boundaries. This paper proposes a target-level incomplete annotation (TIA) for medical image annotation and a multi-lesion segmentation framework. TIA annotates only one complete target region per slice to accurately capture boundaries with minimal annotated effort. Multi-lesion segmentation framework is a weakly supervised learning method, which first implements a medical cut-paste segmentation branch to provide images with pure target pixels and boundaries for training the lesion segmentation model, second utilizes prior anatomical information in the prior-assisted target localization branch to locate and identify target regions, third generates high-confidence pseudo-labels by combining the outputs of cut-paste segmentation branch and prior-assisted target localization branch. A graph neural network (GNN) is adopted to correct noisy labels and propagate reliably labeled pixels to unlabeled pixels. By utilizing TIA, our framework can achieve state-of-the-art results for medical image segmentation, which is validated on Crohn's dataset.