Lesion-Aware Dynamic Kernel for Polyp Segmentation

Automatic and accurate polyp segmentation plays an essential role in early colorectal cancer diagnosis. However, it has always been a challenging task due to 1) the diverse shape, size, brightness other appearance characteristics of polyps, 2) tiny contrast between concealed polyps their surrounding regions. To address these problems, we propose lesion-aware dynamic network (LDNet) for segmentation, which is traditional u-shape encoder-decoder structure incorporated with kernel generation updating scheme. Specifically, designed head conditioned on global context features input image iteratively updated by extracted lesion according predictions. This simple but effective scheme endows our model powerful performance generalization capability. Besides, utilize representation enhance feature background regions tailored cross-attention module (LCA), design efficient self-attention (ESA) capture long-range relations, further improving accuracy. Extensive experiments four public benchmarks collected large-scale dataset demonstrate superior method compared state-of-the-art approaches. The source code available at https://github.com/ReaFly/LDNet .