An Explainable Centralized Deep Learning Model for Gastrointestinal Polyp Segmentation Using the Kvasir-SEG Dataset

Gastrointestinal polyps are well-known precursors to colorectal cancer (CRC), making their accurate detection and segmentation during colonoscopy essential for early diagnosis and cancer prevention. Deep learning–based segmentation models trained on publicly available datasets such as Kvasir-SEG have demonstrated promising performance; however, two key challenges remain: limited robustness across diverse polyp morphologies and endoscopic imaging conditions, and the lack of interpretable decision-making mechanisms that support clinical trust and validation. Many existing centralized segmentation approaches are primarily optimized using overlap-based metrics such as the Dice coefficient and intersection over union (IoU), without adequately analyzing challenging cases such as small, flat, or low-contrast polyps or providing insight into the visual cues influencing model predictions. This study presents an explainable centralized deep learning segmentation model for gastrointestinal polyp segmentation using the Kvasir-SEG dataset. The approach integrates a ResUNet++-Lite encoder–decoder segmentation model with Grad-CAM and masked Grad-CAM visualizations to analyze the spatial regions influencing segmentation predictions. The study focuses on establishing a reproducible and interpretable experimental model that combines systematic preprocessing, data augmentation, centralized training, and explainability analysis. Experimental evaluation on an 80:20 train–test split of the Kvasir-SEG dataset, where data augmentation was applied after splitting, demonstrates stable training behavior and competitive segmentation performance, achieving a pixel accuracy of 0.964, a Dice coefficient of 0.858, and an IoU of 0.791 on the held-out test set. Qualitative explainability results further indicate that the model consistently focuses on anatomically relevant polyp regions. Overall, the study illustrates how segmentation performance and explainable AI techniques can be integrated to support the development of clinically interpretable AI-assisted colonoscopy systems.