Polyp segmentation with convolutional MLP

Yan Jin, Yibiao Hu, Zhiwei Jiang, Qiufu Zheng

Abstract: Accurate polyp segmentation can help doctors find and resect abnormal tissue and decrease the chances of polyps changing into colorectal cancer. The current polyp segmentation neural networks are still challenged by complicated scenarios where polyps have large variations of shapes, size, color, and appearance. In this paper, we propose convolutional multilayer perceptron polyp segmentation network to achieve more accurate polyp segmentation in colonoscopy images. The proposed network adopts a convolutional MLP encoder and enhances the low-level feature using the parallel self-attention module. Furthermore, instead of directly adding encoder features to the decoder, we introduce a cascaded context aggregation module to aggregate the high-level semantic feature and low-level local feature. Finally, channel guide group reverse attention is used to enhance structural and textural details by mining the relationship between areas and boundary cues. The proposed approach is evaluated on six widely adopted datasets and demonstrates superior performance compared to other state-of-the-art models.

External IDs:dblp:journals/vc/JinHJZ23