Adaptive Attention Context Network for Polyp Segmentation

Recently, deep convolutional neural networks provide an effective approach for polyp segmentation in colonoscopy images. However, most existing deep convolutional neural network-based methods tend to focus excessively on deep features, overlooking the shallow features that are crucial for accurately localizing small polyp targets, thus failing to achieve satisfactory segmentation performance. In this paper, we propose a novel Adaptive Attention Context Network (AACNet) that aims to automatically and robustly segment polyp lesions, with a specific emphasis on segmenting small polyp targets. Specifically, we introduce an Adaptive Hybrid Attention Module (AHAM) to replace traditional convolutional operations in the decoding stage. The module consists of a channel attention module and a spatial attention module, utilizing skip connections and upsampled feature maps as inputs. Compared to traditional convolutional operations, the AHAM adaptively selects more optimal segmentation results in both the channel and spatial dimensions, effectively preserving shallow features. To capture richer contextual information, AACNet incorporates a Context Extraction Module, which explores contextual information at each layer through global information compression and local information preservation. The results demonstrate that our proposed method outperforms existing approaches in polyp segmentation, particularly for segmenting small polyp targets.