Retrospective detail reconstruction network for mitigating shallow information loss in colorectal polyp segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Existing polyp segmentation networks struggle with shallow information loss, where low-level features are irreversibly suppressed.
• This issue is particularly problematic for fractal-like polyps, degrading segmentation performance.
• Current methods lack progressive feature guidance, leading to insufficient global semantic awareness.

Purpose of the Study:

• To introduce RDNet, a novel network designed to overcome shallow information loss and enhance polyp segmentation.
• To improve the delineation of polyp boundaries and regions, especially in complex cases.

Main Methods:

• Proposed RDNet, featuring a backward detail reconstruction (BDR) module to recover suppressed low-level cues.
• Introduced a cascaded synergistic optimization (CSO) module for decoupled boundary and region representation with conditional gating.
• Employed layer-wise information injection and cross-layer feature calibration for detail reactivation.

Main Results:

• RDNet consistently outperformed state-of-the-art methods across five benchmark datasets.
• Significant improvements were observed in segmenting fractal-like multi-polyp scenarios.
• The network demonstrated robustness and effectiveness in complex clinical environments.

Conclusions:

• RDNet effectively mitigates shallow information loss in polyp segmentation.
• The proposed BDR and CSO modules enhance structural consistency and semantic discrimination.
• RDNet offers a robust solution for accurate polyp segmentation in challenging medical imaging scenarios.