Perbandingan Strategi Cost-Sensitive Learning Statis dan Dinamis untuk Deteksi Melanoma Menggunakan MobileNetV3-Small pada Dataset HAM10000

Abstract

This study compares the effectiveness of static and dynamic cost-sensitive learning strategies for melanoma detection using the MobileNetV3-Small architecture on the HAM10000 dataset, which has a class imbalance of 1:6 between the melanoma and melanocytic nevus classes. Data imbalance causes the model to be biased toward the majority class, thereby increasing the risk of clinically dangerous false negatives. The methods compared include Class Weight, SMOTE, Weighted Random Sampler (WRS), a combination of SMOTE + Class Weight as a static approach, and Focal Loss as a dynamic approach. All experiments were conducted using identical training configurations to ensure that performance differences were influenced solely by the imbalance handling methods. Evaluation was performed using accuracy, precision, recall, F1-score, AUC-ROC, confusion matrix, and Grad-CAM analysis. The results show that all cost-sensitive learning methods successfully improved melanoma recall compared to the baseline of 0.5676. The combination of SMOTE and Class Weight achieved the highest recall of 0.9099, while SMOTE achieved the best AUC value of 0.9318 with a better balance of performance. Grad-CAM analysis indicates that Focal Loss produces a more calibrated probability distribution, evidenced by the highest probability in false negative cases and the lowest probability in false positive cases. These findings suggest that evaluating medical AI models requires more than just aggregate metrics; it also necessitates analyzing model behavior in critical cases to support interpretability and clinical validity.