Abstract
Colorectal lesion segmentation is central to computer-aided colonoscopy, yet many deep-learning systems terminate at mask generation without providing interpretable evidence or clinically usable decision support. This study proposes XAI-ColoSegRec, a six-stage optimization-guided framework integrating standardized preprocessing, comparative semantic segmentation, visual validation, LIME-based explainability, quantitative morphology analysis, and transparent risk-support recommendation. Kvasir-SEG and CVCClinicDB images were combined, enhanced using CLAHE, resized to 256 × 256 pixels, and used to train UNet, DeepLabV3+, and HRNetV2-Lite under an identical TensorFlow protocol with binary cross-entropy and Dice loss. DeepLabV3+ achieved the strongest performance, with Dice 0.8484, IoU 0.7716, precision 0.8655, recall 0.8644, accuracy 0.9657, and AUC 0.9848. Its superiority was supported by 95% confidence intervals, paired Wilcoxon signed-rank tests, ROC analysis, and the smallest train-validation gap. LIME produced the most lesion-centered explanations for DeepLabV3+, with a Focus Index of 0.4816 ± 0.1254, Stability IoU of 0.5328 ± 0.1767, and explanation runtime of 13.12 ± 6.65 s. Area, perimeter, compactness, and solidity were converted through distribution-derived thresholds into Routine, Moderate, and Critical morphology-based support categories, while Hausdorff distance was used for retrospective boundary auditing. Threshold perturbation and Random Forest, SVM-RBF, ordinal SVR, and voting baselines assessed rule stability and learnability. Module-wise ablation confirmed preprocessing, backbone selection, explainability, and recommendation logic provided complementary, non-substitutable contributions. XAI-ColoSegRec therefore links statistically validated segmentation with auditable explanations and morphology-derived decision support, while requiring external multicentre, clinician-annotated, and histopathologically confirmed validation before diagnostic deployment.