An adaptive fusion of composite attention convolutional neural network for polyp image segmentation

Accurate localization and segmentation of polyp lesions in colonoscopic images are crucial for the early diagnosis of colorectal cancer and treatment planning. However, endoscopic imaging is often affected by noise interference. This includes issues like uneven illumination, mucosal reflections, and motion artifacts. To mitigate the impact of such interference on segmentation performance, it is essential to integrate multi-scale feature analysis effectively. Features at different scales capture distinct aspects of image information. Yet, existing methods typically rely on simple feature summation or concatenation. These methods lack the capability for adaptive fusion across scales.

To address these limitations, this paper proposes AFCNet-an Adaptive Fusion Composite Attention Convolutional Neural Network. AFCNet is designed to improve robustness against noise interference and enhance multi-scale feature fusion for polyp segmentation. The key innovations of AFCNet include: (1) integrating depthwise separable convolution with attention mechanisms in a multi-branch architecture. This allows for the simultaneous extraction of fine details and salient features. (2) Constructing a dynamic multi-scale feature pyramid with learnable weight allocation for adaptive cross-scale fusion.

Extensive experiments on five public datasets (ClinicDB, Kvasir-SEG, etc.) demonstrate that AFCNet achieves state-of-the-art performance, with improvements of up to 3.73 $\%$ in Dice coefficient and 4.62 $\%$ in IoU, validating its effectiveness and generalization capability in polyp segmentation tasks.

AFCNet is a novel polyp segmentation network that leverages convolutional attention and adaptive multi-scale feature fusion, delivering exceptional generalization and adaptability.