Symmetry-Localized Non-Local Blocks for Polyp Recognition and Segmentation on Endoscopic Images

This work is dedicated to introducing specialized self-attention blocks as a part of deep neural network-based models in polyp recognition and segmentation using images obtained through video endoscopy. In our study we introduce several modifications of non-local blocks that implement the concept of self-attention, considering the specificity of biomedical data. Using the changes proposed, we improved the performance of the considered deep neural network baseline (from 0.710 to 0.744 F1-score for the classification problem statement and from 0.805 to 0.823 Intersection over Union for the segmentation problem statement) on joined test subsamples of publicly available datasets containing annotated endoscopy images for polyps recognition.