TITLE:
A Novel Polyp Segmentation Method Based on the Vision Transformer and Attention Mechanism
AUTHORS:
Xinping Guo, Yongqi Nie, Xiuzhu Jia, Mengying Lou, Zhiyuan Li, Xiaoyu Han, Lu Yu
KEYWORDS:
Polyp Segmentation, UNet Model, The Attention Mechanism, The Pyramid Vision Transformer
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.6,
June
23,
2026
ABSTRACT: The accurate segmentation of the polyp is very important for the diagnosis and treatment plans of colorectal cancer. Although the UNet model and models with the U-shaped structure have achieved great success in polyp image segmentation, they are still limited by the colors, sizes, and shapes of polyps, as well as the low contrast, various noise, and blurred edges of the colonoscopy, which can easily result in a large amount of redundant information, weak complementarity between different levels of features, and inaccurate polyp localization. To deal with the special characteristics of the polyp images and improve the segmentation performance, a new segmentation model named VTANet, which is based on the pyramid vision transformer and BAM (Bottleneck Attention Module), is developed. The proposed model consists of four modules: the pyramid vision transformer (PVT) encoder, the Feature Aggregation Module (FAM), the Adaptive Attention Fusion Module (AAFM), and the Aggregation Similarity Module (ASM). The PVT learns a more robust representation model; the FAM enhances the complementarity between features by cascading the encoder features and acquiring richer context and fine-grained features. The AAFM makes polyp localization more accurate by introducing the BAM attention module to obtain richer details of the polyps. To verify effectiveness and accuracy, experiments on five popularly used datasets are carefully designed and implemented. The proposed VTANet achieves competitive and generally superior performance across five public datasets. Although it does not obtain the best score on every metric, especially on several MAE or E-measure results, it consistently improves the main overlap-based metrics such as mDice and mIoU on most datasets. This indicates that VTANet provides a favorable balance between region-level accuracy, boundary preservation, and generalization ability.