A Unified Framework for Classification and Segmentation of Ambiguous Dual-Type Lesions in Colonoscopic Images.

Accurate analysis of lesions in colonoscopic images is essential for computer-aided diagnosis. However, most existing methods are designed for single-lesion segmentation and assume a predefined lesion category, limiting their applicability in real-world scenarios where multiple lesion types exhibit similar visual characteristics. To address this issue, we propose a unified framework for the joint classification and segmentation of dual-type lesions in colonoscopic images, enabling simultaneous identification and localization of submucosal lesions and polyps/adenomas. The proposed method integrates joint supervision, context-aware feature enhancement, and ambiguity-aware optimization to improve consistency between semantic recognition and spatial delineation. In particular, a soft-label supervision strategy is introduced to alleviate semantic ambiguity, while an imbalance-aware loss design enhances segmentation accuracy and reduces false negative predictions. Extensive experiments on both private and public datasets demonstrate that the proposed method achieves superior performance compared with representative CNN- and transformer-based approaches. Notably, the method shows clear advantages in segmentation accuracy, localization precision, and robustness under challenging conditions. Ablation studies further confirm the effectiveness of each component in the proposed framework. These results indicate that the proposed approach provides an effective solution for dual-type lesion analysis and has the potential to assist clinical decision-making in gastrointestinal endoscopy.