XPolypNet: A U-Net-Based Model for Semantic Segmentation of Gastrointestinal Polyps With Explainable AI

Abstract

Automated segmentation of gastrointestinal polyps is a critical step in the early detection and prevention of colorectal cancer (CRC), which is one of the most common causes of cancer-related deaths worldwide. This article presents a U-Net-based model enhanced with Attention Mechanisms and Atrous Spatial Pyramid Pooling (ASPP) for accurate polyp segmentation. To address the challenges of varying polyp sizes, indistinct boundaries, and complex textures, the model used a combined loss function (Binary Cross-Entropy and Dice Loss). Additionally, Gradient-Weighted Class Activation Mapping (Grad-CAM) was integrated to provide visual explanations of the model’s decisions to increase trust and interpretability by clinical practitioners. The presented model was evaluated on five benchmark datasets, achieving a Dice Coefficient of 0.8378 and a Mean Intersection over Union (mIoU) of 0.8427. The comparative analysis highlighted its superiority when compared to state-of-the-art contemporary approaches, with a precision and accuracy of 97%. Qualitative analyses also underline the ability to accurately delineate polyps, even in difficult situations. Although the model exhibited satisfactory performance, it still faced challenges regarding boundary misclassification and reduced efficacy in datasets with high variability. The next steps of this research will focus on domain adaptation and integration of additional modalities to enhance generalizability. This study provides a step toward automated polyp detection and demonstrates the potential of explainable artificial intelligence (XAI) to change the accuracy of diagnosis and healthcare for patients.