Learning Pixel Level Affinity with Class Labels for Weakly Supervised Segmentation of Lung Cavities

Accurately annotating lung cavities (LCs) at the pixel level in computed tomography (CT) images presents a significant challenge due to their diverse shapes and sizes. To address this limitation, weakly supervised semantic segmentation (WSSS) methods utilizing sparse annotations, such as image-level labels, have emerged as a promising trend. This paper proposes a novel scribble-supervised segmentation framework for LCs that leverages image-level annotation-driven pixel affinity. The framework introduces a bidirectional interaction Mamba UNet model, named MambaUNeLCsT, designed to address the inefficiency of transformer models in processing long sequences. To refine coarse LCs pseudo-labels, an attention-based affinity pseudo-label refinement module is incorporated, employing an affinity algorithm to establish associations between unlabeled and pseudo-labeled samples. This approach infers labels for unlabeled samples by computing sample similarities. Additionally, to overcome the limited spatial supervision provided by image-level annotations, a scribble-based segmentation module is included, effectively capturing the complete morphology and boundary information of LCs. This module enhances the model’s capability to recognize and process fine structures. Experimental results demonstrate that MambaUNeLCsT achieves state-of-the-art performance in 3D medical image segmentation, outperforming existing models in WSSS tasks.