MECSNet: a dual-encoder multi-scale attention network for polyp segmentation

Résumé fourni par la source

Abstract Polyp segmentation in medical imaging imposes stringent requirements on the feature extraction capability and robustness of the model, while traditional convolutional networks are deficient in global information capture, boundary blurring, and multi-scale modeling. To address these issues, this paper proposes a segmentation framework, MECSNet, based on a dual encoder-decoder architecture of convolutional neural network (CNN) and context-aware transformer (CAFormer). The core innovation of MECSNet is the design of a multi-branch augmented convolutional block MECS, which effectively improves the feature identification capability and robustness of the model by combining the average-pooling, maximal-pooling, and median-pooling strategies with the multiscale deep convolution, which also improves the feature expression ability and anti-noise performance. To enhance the model’s ability to perceive small targets and boundaries, the coordinate attention (CoordAtt) spatial attention module is introduced in the jump connection to enhance spatial information modeling. Experiments on several mainstream polyp segmentation benchmark datasets show the superior performance of MECSNet, especially on the Kvasir-SEG dataset, where the average Dice coefficient and IoU values are significantly improved by 1.3% and 1.8%, respectively, compared with existing state-of-the-art methods.