<![CDATA[Deep learning models for dermoscopic skin tumor classification typically default to a static 0.5 probability threshold. In highly imbalanced clinical datasets, this rigid boundary yields unacceptable false-negative rates for critical malignancies. This study proposes a lean diagnostic framework designed to maximize clinical triage safety and computational efficiency for edge-based environments. An EfficientNet-B3 architecture was trained on ISIC imagery using an Alpha-Weighted Binary Focal Loss. To minimize computational overhead, we applied true structural channel pruning via L2-norm magnitude at 0%, 30%, and 60% sparsity. Rather than relying on default cutoffs, we dynamically calibrated the decision boundary by applying Youden’s J statistic to the Receiver Operating Characteristic (ROC) validation curve to maximize discriminative sensitivity. Network surgery at 60% sparsity reduced parameters from 10.70M to 1.81M and FLOPs from 1.93G to 0.37G. By recalculating predictions with the Youden-derived optimal threshold (0.4110), the lightweight model sustained a high malignant recall (sensitivity) of 82.54% and an Area Under the Curve (AUC) of 0.8457, closely rivaling the unpruned baseline recall of 84.72%. Grad-CAM evaluations confirmed that despite the physical severing of 60% of network channels, the model preserved high spatial fidelity in localizing relevant morphological features. Shifting from static probability thresholds to statistically optimized boundaries allows highly pruned, computationally efficient networks to maintain safe triage recall rates, providing a robust framework for automated, resource-constrained skin cancer screening.]]>
Copyrights © 2026
