Endoscopy-based lesion segmentation is essential for the early detection of neoplastic lesions. Framed as a dense-prediction task on high-resolution mucosal panoramas, its automated and precise contouring alleviates clinicians’ real-time cognitive burden, while yielding quantitative biomarkers to stratify patient risk and guide subsequent therapeutic strategies. However, three key challenges persist in current methods: (1) multi-scale confusion due to diverse lesion morphologies that resemble surrounding tissues; (2) weak boundaries that hinder foreground-background distinction; (3) over-reliance on training data that limits cross-dataset generalization. We propose MFG-Seg, a Multi-Level Feature Guide Segmentation Framework with a dual-decoder architecture. We design a Boundary-Aware (BA) decoder as the primary decoder, which effectively mitigates the loss of boundary details caused by deep encoder layers. In addition, we introduce a Multi-scale Selective Attention (MSA) decoder as the auxiliary decoder, which augments discrimination of confusing textures via multi-scale fine-grained features and cross-attention. Extensive experiments on six datasets validate MFG-Seg’s superior performance, achieving a lightweight design with only 25.39M parameters and 7.19 GFLOPs. Compared with existing methods, MFG-Seg demonstrates significant improvements in generalization accuracy and clinical deployability for endoscopic images diagnosis.
Xu et al. (Sat,) studied this question.