Hybrid CNN Architecture for Skin Cancer Classification

This thesis investigates automatic skin cancer classification by integrating dermoscopic images and structured clinical metadata from the ISIC 2024 dataset (≈401k images, 55 variables). A hybrid architecture is proposed, combining a Convolutional Neural Network (CNN) for image analysis with multiple learners on metadata, including a Feedforward Neural Network (FFNN), Random Forest, and Gradient Boosting. The outputs of these base models are aggregated through a logistic regression stacking meta-model, leveraging out-of-fold predictions to prevent data leakage. The evaluation, performed on a 25% hold-out set with extremely low prevalence of positives (≈0.1%), highlights the advantages of model hybridization. The stacked meta-model achieves superior performance compared to individual learners (ROC-AUC 0.953), while Precision-Recall analysis (AP ≈0.095) reflects the class imbalance. The optimal decision threshold, chosen using Youden’s index (t* ≈0.001), prioritizes recall (0.867) at the cost of precision (≈1%), a suitable strategy in triage contexts where minimizing false negatives is critical. The work underscores the importance of ensemble learning, cost-sensitive metrics, and uncertainty calibration for clinical deployment. Future directions include imbalance-aware loss functions, advanced CNN backbones and context-aware augmentation. Overall, this study contributes to the development of interpretable and regulation-aligned AI systems for dermatological decision support.