Multi-layer Feature Fusion and Coarse-to-fine Label Learning for Semi-supervised Lesion Segmentation of Lung Cancer

Jiale Chen, Siyang Feng, Yanfen Cui, Chuansong Fan, Huan Lin, Xinjun Bian, Lingqiao Li, Zhenbing Liu, Zaiyi Liu, Rushi Lan, Xipeng Pan

Abstract: Lung lesion segmentation is essential for medical diagnosis, and prognosis evaluation. However, the significant expense of pixel-level annotations limits the availability of labeled data and brings challenges in precise segmentation. Semi-supervised learning (SSL) presents a promising alternative by leveraging a large volume of unlabeled samples to reduce dependence on labeled data. Although previous works had made some progresses, SSL still faces two major challenges: (1) inadequate robustness of consistency learning in ambiguous regions, and (2) inefficient utilization of information from noisy pseudo-labels. To overcome these challenges, we propose a novel SSL segmentation framework, named Multi-layer Feature Fusion and Coarse-to-fine Label Learning for Semi-supervised Lesion Segmentation of Lung Cancer (MCL3S). The framework incorporates Multi-layer Feature Fusion (MFF), which introduces random perturbations into features during the embedding-to-pixel transition, enhancing segmentation robustness in ambiguous regions. Furthermore, we propose two additional modules: Pixel-level Pattern Relocalization Module (PPRM) to capture fine-grained local contextual information, and Ambiguity Perceptual Learning (APL) to extract valuable insights from noisy pseudo-labels by focusing on uncertainty regions. Experimental results demonstrate that MCL3S achieves significant performance improvements over seven state-of-the-art semi-supervised segmentation approaches on two in-house lung cancer datasets and one public dataset. The implementation code will be publicly available at https://github.com/Pahoia/MCL3S.