Implementation of MobileNetV2-Based Convolutional Neural Network for Melanoma Detection Using Dermoscopic Images

Abstract

Melanoma is an aggressive form of skin cancer, making early detection essential for the timely identification of potentially malignant skin lesions and the improvement of diagnostic outcomes. This study aims to evaluate a convolutional neural network (CNN) based on the MobileNetV2 architecture for classifying dermoscopic images into benign and malignant categories using the HAM10000 dataset. Image preprocessing involves resizing and normalization, while additional techniques, including contrast-limited adaptive histogram equalization (CLAHE), hair removal, and soft segmentation, are applied to enhance image quality, reduce visual artifacts, and improve lesion visibility. The model is developed using transfer learning with weights pretrained on ImageNet, followed by training of the classification head and fine-tuning of the final network layers to adapt the extracted features to the target classification task. Model performance is assessed using accuracy, loss, precision, recall, and F1-score to provide a comprehensive evaluation of classification performance. Evaluation on a primary test set comprising 659 images yields an accuracy of 85.13%, a loss of 0.3430, and a sensitivity of 88.33% for the malignant class. A separate evaluation conducted on 128 images using Google Colab achieves an accuracy of 87.50%; these results are reported independently of the primary evaluation. The findings indicate that MobileNetV2 has potential for dermoscopic skin lesion classification, particularly in identifying malignant lesions. However, further evaluation using independent datasets, cross-dataset validation, and computational efficiency analysis is necessary to assess the model's generalizability and reliability before considering its potential application in clinical settings.