PRCNet: A parallel reverse convolutional attention network for colorectal polyp segmentation

Colorectal polyps are one of the common gastrointestinal disorders. The precise segmentation of these polyps will assist physicians in diagnosing and treating these lesions. Polyps typically exhibit a variety of sizes and shapes, coupled with diverse scale variations. Additionally, the boundary between the polyp and the adjacent mucosa often lacks a clear definition. This makes segmenting polyps a challenging task in clinical practice. Although deep learning-based methods have already been employed to tackle this issue, especially in the face of challenging environmental segmentation difficulties, many of these approaches fall short. They often lack features that aggregate multi-scale and multi-receptive field characteristics, neglecting regional boundary constraint considerations. To address these limitations, we propose a parallel fusion feature colorectal polyp segmentation network (PRCNet) that aims to achieve precise polyp segmentation in colonoscopy images. Specifically, PRCNet involves a convolutional block attention module (CBAM), a reverse attention module (RA), a parallel fusion decoder (PFD), and a receptive field module (RFB). The PFD integrates neighboring features across various levels and aggregates high-level feature layers. This approach allows the exploration of information at multiple levels. Conversely, the RA and CBAM have been developed to effectively capture synergies among regions, boundaries, channels, and spaces, thereby enhancing the accuracy and robustness of segmentation. The empirical study implements PRCNet on five challenging datasets, outperforming mainstream medical segmentation models. The results consistently show that PRCNet surpasses PraNet, a widely used network in this field. Specifically, we train PRCNet on several datasets and observe marked improvements, including a 2.6% increase on the CVC-ClinicDB, 9.8% on the ETIS dataset, and 1.4% on the Kvasir dataset. Additionally, PRCNet demonstrates significant advancements in its generalization capacity.