Abstract
Colorectal cancer (CRC) continues to be a significant global health issue as a result of the difficulties associated with the early and precise detection of lesions through colonoscopy. We propose CADxPolyDetect, a clinically explainable hybrid deep learning system for multi-class colorectal lesion detection using augmented colonoscopy images, to address diagnostic limitations that arise from class imbalance, interpretability gaps, and spatial complexity in colonoscopic images. A five-stage experimental pipeline is integrated into the framework. Initially, the original Hyper Kvasir dataset (10,672 images) is augmented to 23,000 images using Deep Convolutional Generative Adversarial Networks (DCGAN) and class balanced using SMOTE. Three pre-trained CNNs—ResNet-50, DenseNet-201, and VGG-16—are implemented in Stage 2 to execute end-to-end feature extraction. Transformer networks are integrated with these CNNs in Stage 3 to capture global spatial dependencies. Stage 4 incorporates multi-class support vector machines (SVM) for the final classification of 23 colorectal conditions. Stage 5 improves interpretability by utilising Grad-CAM to produce clinically relevant heatmaps that localise malignant polyps and facilitate transparent decision-making. The DenseNet-201 + Transformer + Multi-Class SVM model showed the best results of all the setups tested, achieving an overall accuracy of 98%, an F1-score of 0.98, a precision of 97%, and a recall of 99%, based on a 70:30 split of training and testing data. Robustness, precision, and clinical trust are guaranteed through the integration of attention-based visualisations, deep ensemble learning, and augmented data. Therefore, CADxPolyDetect sets a new benchmark in AI-supported colonoscopy by offering a decision support system that is easy to understand and can grow to meet needs for the early detection and screening of colorectal cancer.