GAML: Geometry-Aware Mutual Learning for polyp segmentation

doi:10.1016/j.bspc.2025.107965

GAML: Geometry-Aware Mutual Learning for polyp segmentation

2025-05-29

0

PRE

AI

DOI：10.1016/j.bspc.2025.107965

Original

Original request for help

Save

Abstract

En 中文

Accurate and automatic polyp segmentation plays a pivotal role in early colorectal cancer diagnosis. However, this task remains challenging due to (i) the diverse sizes and irregular shapes of colonic polyps, and (ii) the ambiguity in polyp boundaries with the surrounding mucosa. To overcome these challenges, we propose a novel and effective Geometry-Aware Mutual Learning (GAML) for precise polyp segmentation, incorporating the following steps. First, we design a region feature and boundary-wise extractor to mine the internal and external geometric features of the target. Secondly, we introduce a geometry-aware context aggregation module (GACA) to learn long-range dependencies between pixels in the external structure region and pixels within the object. Thirdly, we present a geometry-aware graph reasoning module (GAGR) that enhances and propagates features across different domains. This is accomplished by utilizing region and boundary features as nodes in the graph, leading to accurate segmentation along the boundaries. Finally, the network incorporates latent representations of internal and external geometric structures and is adaptively trained to recognize and segment polyp lesions. Experimental results on four benchmark datasets (e.g., Kvasir-SEG, CVC-ClinicDB, BKAI, and Kvasir-Sessile) demonstrate that our method outperforms current state-of-the-art methods regarding various evaluation metrics. The source code is available at https://github.com/DLWK/GAML.

Keywords:

Polyp segmentation

Geometry-Aware

Mutual learning

Computer-aided diagnosis (CAD)

GAML: Geometry-Aware Mutual Learning for polyp segmentation

Abstract

Journal

Organization