ColoXAI-RecomNet: Explainable Recommender Framework for Colorectal Cancer Classification Using Integrated CNN...

The classification of colorectal disease based on colonoscopy images requires not only high predictive accuracy but also interpretable decision support. This study proposes a five-stage explainable framework for multi-class colorectal image classification on the Kvasir dataset. This framework is named ColoXAI-RecomNet. The initial stage of this framework involved an investigation of three parallel hybrid CNN ensemble models: RDV-2025 (ResNet50 + DenseNet121 + VGG16), IEM-2025 (InceptionV3 + EfficientNetB0 + MobileNetV2), and DRE-2025 (DenseNet201 + ResNet101 + EfficientNetB3). These CNN models were used as parallel feature extractors. Their combined deep features were used for classification. Among these models, DRE-2025 was identified as one of the top performers in Stage 1. This model had an approximate scale of 77.2 M for its backbones. The enhanced features for decision were passed to a multi-class SVM for final classification in Stage 3. The CNN features were refined by an XGBoost model in Stage 2. The final output of this framework was converted to a clinically interpretable recommender output in Stage 5. LIME was incorporated into this framework in Stage 4 to provide visual explanations for each image. The DRE-2025 + XGBoost + SVM model, which performed best, achieved an accuracy of 98.60%, a precision of 98.75%, a recall of 98.50%, an F1-score of 98.62%, and an AUC of 99.30%. In Stage 2, the combined CNN features were improved using an XGBoost algorithm. Stage 3 then classified the Stage 2 outputs using a multi-class SVM with an RBF kernel. Stage 4 used LIME for visual explanation, and Stage 5 converted the final prediction to a clinically interpretable recommender output.