Skin cancer classification using a borderline-SMOTE enhanced neural network model on dermoscopic images

초록

Skin cancer is one of the most common and life-threatening cancers worldwide, with high mortality rates when diagnosis is delayed. Accurate classification of skin cancer using dermoscopic images is challenging due to inconsistencies in lighting, skin texture, lesion shape, and imbalanced datasets. This study presents an automated method for multiclass skin cancer classification to assist dermatologists in early detection. We introduce a deep learning model based on the enhanced MobileNetV2 architecture, designed to classify multiple types of skin lesions. To address dataset imbalance and reduce overfitting, the borderline-SMOTE technique is employed to oversample underrepresented classes. The model is trained and validated using the ISIC 2020 dataset, which comprises dermoscopic images collected from diverse clinical sources. The performance of the proposed model is compared against three widely used convolutional neural networks: InceptionV3, EfficientNet, and ResNet50. The proposed model achieved 98.50% accuracy, 98.50% precision and recall, and an F1-score of 98.47%. It also recorded an Area Under the Curve (AUC) of 99.63%, outperforming all baseline models evaluated on the same dataset. These results demonstrate the effectiveness of the proposed borderline-SMOTE enhanced MobileNetV2 model in multiclass skin cancer classification. Its high accuracy and lightweight design make it well-suited for integration into clinical decision support systems, with the potential to improve early diagnosis and patient outcomes.

키워드