PPFormer: A Novel Model for Polyp Segmentation in Digestive Endoscopy

Polyp segmentation is a pivotal task in the field of medical image processing. We devised a more effective deep learning model (PPFormer) that seamlessly integrates pyramid pooling module with transformer. This integration significantly improves the model’s ability to restore intricate details during the decoding phase. Additionally, we rethinked the importance of multi-scale feature maps within the model and thoughtfully designed two pruning strategies to target the elimination of redundant and mis-segmented feature maps, resulting in improved segmentation quality. In this paper, we aim to explore methods to enhance the performance of the polyp segmentation model. We conducted experiments on three different polyp segmentation datasets, and the model presented in this paper consistently exhibited exceptional performance. Through visual experiments, the model demonstrated an enhanced capacity to handle the edge of the polyp, indicating an improved capability to restore image details during the decoding process. In terms of quantitative metrics, PPFormer achieved outstanding results in segmentation-related indicators. For example, it obtained mIoU scores of 91.67%, 92.09%, and 93.19% on the Kvasir-SEG, CVC-ClinicDB, and CVC-300 datasets, respectively.