Improved Colorectal Polyp Segmentation Using Enhanced MA-NET and Modified Mix-ViT Transformer

Abstract

Colorectal polyps is a prevalent medical condition that could lead to colorectal cancer, a leading cause of cancer-related mortality globally, if left undiagnosed. Colonoscopy remains the gold standard for detection and diagnosis of colorectal neoplasia; however, a significant proportion of neoplastic lesions are missed during routine examinations, particularly diminutive and flat lesions. Deep learning techniques have been employed to improve polyp detection rates in colonoscopy images and have proven successful in reducing the miss rate. However, accurate segmentation of small and flat polyps remains a major challenge to existing models as they struggle to differentiate polypoid and non-polypoid regions apart. To address this issue, we present an enhanced version of the Multi-Scale Attention Network (MA-NET) that incorporates a modified Mix-ViT transformer as the feature extractor. The modified Mix-ViT facilitates ultra-fine-grained visual categorization to improve the segmentation accuracy of polypoid and non-polypoid regions. Additionally, we introduce a pre-processing layer that performs histogram equalization on input images in the CIEL*A*B* color space to enhance their features. Our model was trained on a combined dataset comprising Kvasir-SEG and CVC-ColonDB and cross-validated on CVC-ClinicDB and ETIS-LaribDB. The proposed method demonstrates superior performance compared to existing methods, particularly in the detection of small and flat polyps.