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

Area of Science:

• Gastrointestinal endoscopy imaging analysis
• Medical image analysis and computer-aided diagnosis
• Computational pathology and oncology

Background:

• Accurate colonoscopic image analysis is crucial for computer-aided diagnosis.
• Existing methods often focus on single lesion types, limiting real-world application.
• Multiple lesion types can present similar visual characteristics, complicating analysis.

Purpose of the Study:

• To develop a unified framework for joint classification and segmentation of dual-type lesions in colonoscopic images.
• To enable simultaneous identification and localization of submucosal lesions and polyps/adenomas.
• To improve consistency between semantic recognition and spatial delineation of lesions.

Main Methods:

• Integration of joint supervision, context-aware feature enhancement, and ambiguity-aware optimization.
• Implementation of a soft-label supervision strategy to mitigate semantic ambiguity.
• Design of an imbalance-aware loss function to boost segmentation accuracy and reduce false negatives.

Main Results:

• The proposed method demonstrates superior performance over existing CNN- and transformer-based approaches on public and private datasets.
• Significant advantages observed in segmentation accuracy, localization precision, and robustness under challenging conditions.
• Ablation studies validate the effectiveness of individual framework components.

Conclusions:

• The developed approach offers an effective solution for dual-type lesion analysis in colonoscopy.
• This method has the potential to significantly aid clinical decision-making in gastrointestinal endoscopy.
• The framework improves the accuracy and reliability of lesion detection and characterization.