Shallow Attention Network for Polyp Segmentation

Accurate polyp segmentation is of great importance for colorectal cancer diagnosis. However, even with a powerful deep neural network, there still exists three big challenges that impede the development segmentation. (i) Samples collected under different conditions show inconsistent colors, causing feature distribution gap and overfitting issue; (ii) Due to repeated downsampling, small polyps are easily degraded; (iii) Foreground background pixels imbalanced, leading biased training. To address above issues, we propose Shallow Attention Network (SANet) Specifically, eliminate effects color, design color exchange operation decouple image contents force model focus more on target shape structure. Furthermore, enhance quality polyps, shallow attention module filter out noise features. Thanks high resolution features, can be preserved correctly. In addition, ease severe pixel imbalance probability correction strategy (PCS) during inference phase. Note though PCS not involved in training phase, it work well consistently improve performance. Quantitative qualitative experimental results five challenging benchmarks confirm our proposed SANet outperforms previous state-of-the-art methods by large margin achieves speed about 72FPS.