Machine Learning Based Classification of Medical Images of Melanoma Using Transfer Based Learning

Abstract

Over 7650 people died from the United States died from melanoma just within 2022, with this number projected to increase by 4.4% in 2023 according to the Skin Cancer Foundation. By accurately diagnosing this disease and implementing an accurate model into the healthcare space for clinical decision-making, early classification of skin lesions may increase the likelihood of treatment before cancer metastasis. Machine learning diagnosis has gained attention in the previous years and has been proven to contribute to the early diagnosis of various diseases. In the context of skin cancer diagnosis, there is a limited amount of medical images, making it challenging to utilize typical machine learning approaches for classification. Therefore, in this work, we utilize transfer learning for the automated classification of medical images of melanoma into benign and malignant. Accordingly, we develop a transfer based algorithm based on a pre-trained InceptionV3 and a VGG16 model in Python. We compared the performance of these two models in order to evaluate the optimal model. The number of epochs and the learning rate were optimized for both models. In order to assess the model, we utilize a variety of metrics, including confusion matrix, ROC, AUC, accuracy, sensitivity, specificity, precision, and F1 score. The results of this study demonstrated that an optimized VGG16 model outperformed VGG16 and was able to successfully classify at least 90% of testing images. The developed model can potentially contribute to the early and automated diagnosis of melanoma in clinical settings.