Title: Application of Deep Learning Model in Skin Lesion Classification
Authors: Hoang Quan Nguyen , Minh Duc Nguyen
Volume: 9
Issue: 12
Pages: 13-21
Publication Date: 2025/12/28
Abstract:
Considering the differential diagnosis, skin cancer, and especially, malignant melanoma, is often a diagnosis of a challenge owing to the case and expertise of the clinician, and serious inter-observer variability. These scenarios in the absence of a dermatologist, often leads to missed or delayed diagnosis. Thus, the support for deep learning based on clinical decision-making has emerged as an important aid in dermatological screening. This work analyses deep learning on skin lesion classification and inference on the clinically motivated level of the lesion using ResNet50 and MobileNetV2, two convolutional neural networks. The experiments carried on were HAM10000 dataset, containing the seven dermoscopic skin lesions fully imbalanced in the size of each category. Given the common methodological flaws, there was lesion-ID based data splitting to prevent leakage of data along with an imbalance aware training using controlled oversampling and focal loss. The performance of the models was evaluated with a mix of comprehensive set of evaluation metrics and using confusion matrices per level of lesion. ResNet50 demonstrated higher overall accuracy but exhibited a higher rate of false negatives for melanoma. For MobileNetV2, on the other hand, is more useful for the low-resource-setting especially for the less overall accuracy and more oriented error profile concerning sensitivity. Overall, the findings highlight the importance of lesion-level evaluation and clinically aligned performance metrics for developing reliable AI-based skin cancer screening systems.