Melanoma Detection System Using Deep Learning Techniques

Abstract

Melanoma is a very aggressive form of skin cancer. Diagnosis by dermatologists is based on a thorough clinical examination, dermatoscopy and, if necessary, a biopsy with histopathological analysis. Early diagnosis is crucial for effective treatment and a favourable prognosis. A computer-assisted diagnostic procedure is more objective and reliable than expert human diagnosis, which is subjective and not necessarily reproducible. Machine learning and image processing techniques have been widely used for the automated diagnosis of cutaneous melanoma. More recently, deep learning techniques are being used more and more and are giving better results, especially with convolutional neural network models. The rarity of melanoma cases and concerns about data privacy are major obstacles to creating balanced datasets for the automatic diagnosis of melanoma. In this work, we propose a melanoma detection system based on deep learning techniques. In this system, we address the issue of dataset imbalance in skin data. To this end, we propose using a rebalancing technique by subdividing the database into balanced subsets to reduce bias towards the non-melanoma class. Then, we leverage transfer learning, using the ResNet50 model as the base model. In the final phase of the proposed system, which involves classification and decisionmaking, we tested several machine learning techniques, namely SVM, XG-Boost, logistic regression, and voting techniques: the average voting method and the weighted voting method. Each model was trained and evaluated separately, achieving an accuracy of 81%. After combining the decisions, the best results were obtained through weighted voting, with 87.73%, 86.07%, 90%, and 88.01% for precision, accuracy, recall, and F1 score, respectively. Based on these results, the weighted voting technique balanced the predictions and significantly reduced biases. The system proved to be more reliable and accurate, capable of providing effective computer-aided diagnosis in medical practice.