Identificação de melanomas utilizando redes neurais convolucionais

Abstract

This study aimed to develop and evaluate a model for the automated classification of melanoma and other skin lesions using dermoscopic images. The EfficientNet-B3 architecture with transfer learning was trained on the public HAM10000 dataset, which contains 10,015 images distributed across seven skin-lesion classes. To address class imbalance, the selected configuration included class-weighted Focal Loss, mixup, dropout, weighted sampling, and progressive fine-tuning. The processing workflow also included color-marker and hair removal, contrast enhancement, resizing, and image normalization. On the validation set, the selected configuration achieved an accuracy of 82.38%, a weighted F1-score of 0.828, and a macro F1-score of 0.708. For melanoma, precision was 0.565, recall was 0.609, and the F1-score was 0.586. The results indicate discrimination capability across the seven evaluated categories; however, the moderate melanoma performance and the absence of external validation limit the generalizability of the findings. Therefore, the model should be regarded as an experimental computational prototype for skin-lesion classification, without inference of clinical use at this stage.