<![CDATA[Accurate differentiation between primary and metastatic melanoma in histopathological assessment is critical for staging and therapeutic decision-making. Although deep learning models often report high classification accuracy, their robustness and threshold-dependent clinical behavior remain insufficiently examined. We propose a cross-validated and threshold-optimized deep learning framework for classifying 206 histopathological regions of interest (ROIs), partitioned in an 80:20 split into training (n = 164) and evaluation (n = 42) subsets, using a ResNet-18 backbone. On the hold-out evaluation set, the model achieved an AUC of 0.922. To evaluate generalization stability, stratified 5-fold cross-validation was conducted across all ROIs, yielding fold AUCs ranging from 0.904 to 0.973 and a mean AUC of 0.938 ± 0.024, with a pooled out-of-fold AUC of 0.916. At a decision threshold of 0.5, the model achieved 78.6% accuracy (macro F1 = 0.7846). Increasing the threshold to 0.8 improved accuracy to 85.7% (macro F1 = 0.856), accompanied by higher precision for metastatic melanoma (0.894) and improved recall for primary melanoma (0.904), underscoring clinically meaningful sensitivity–specificity trade-offs beyond AUC alone. Grad-CAM analysis demonstrated spatially coherent activation concentrated within tumor-dense regions in true positives, minimal activation in true negatives, and intermediate activation in a borderline false negative case (probability = 0.75), linking prediction confidence to morphologically relevant evidence. Collectively, these findings highlight the importance of cross-validation rigor, threshold calibration, and interpretability in advancing clinically reliable deep learning systems for melanoma classification.]]>
