A Real-Time Enhanced Channel Attention Network for Endoscopic Polyp Segmentation

Accurate segmentation of polyps during endoscopy is crucial for the correct diagnosis of gastrointestinal diseases. With advancements in deep learning, polyp segmentation models have made significant progress. However, their deployment on medical devices is challenging due to their high computational requirements. To address this challenge, a real-time enhanced channel attention network is proposed for segmenting endoscopic polyps. Our method first introduces a feature extraction block based on the inverted residual block to preliminarily extract image features. Following this, an enhanced channel attention block is introduced, which incorporates standard deviation pooling to enhance sensitivity to polyp textures and boundaries, strategically extracting features for polyp segmentation. Then segmentation map is achieved through a feature decoder. During the training process, a hybrid loss function combining cross-entropy and Dice coefficients is adopted, which is particularly effective for polyp segmentation scenarios. Extensive experiments on two benchmarks including Kvasir and PolypGen demonstrate that our method achieves state-of-the-art performance and runs at 110.73 FPS with only a few parameters (4.94M) are introduced, which confirms its effectiveness and efficiency.