PolyPooling: An accurate polyp segmentation from colonoscopy images

Abstract

The automatic and precise segmentation of polyps in colonoscopy images has gained significant attention in recent artificial intelligence and computer vision research communities. Early and accurate segmentation provides valuable information for medical treatment. Transformer-based models have been widely adopted in this context. Based on the changes in the attention-based module, the emergence of PoolFormer has demonstrated its feasibility, producing promising results across various computer vision tasks. In this paper, PolyPooling is proposed. This model is based on an encoder–decoder architecture specifically designed for the accurate segmentation of polyps in endoscopic images. The encoder is developed based on the PoolFormer architecture, while the decoder utilizes a hierarchical structure to decode multi-level features with support from the Convolutional Block Attention Module (CBAM) and Hamburger module. To further enhance the refinement of polyp boundaries in the global map, a new refinement module is given that replaces traditional attention-based modules with the pooling operator while keeping a residual connection. Experimental evaluations conducted on five challenging and up-to-date datasets demonstrate that PolyPooling outperforms many state-of-the-art methods, especially achieving an improvement of nearly 12.1% in mDice and 12.7% in mIoU compared to the latest method on the ETIS-Labrib dataset. On the same metrics, PolyPooling demonstrates notable performance on other datasets, achieving an improvement ranging from 1.8% to 4.4% on the CVC-ColonDB and CVC-T datasets. It remains competitive on the Kvasir and CVC-ClinicDB datasets. Moreover, PolyPooling fixes the top performance curve among all the methods in F1-score within Kvasir and CVC-ClinicDB datasets. Our implementation code is available at: https://github.com/long-nguyen12/PolyPooling.