A Comparison of Convolutional Neural Networks used in Melanoma Detection : With transfer learning on the PAD-UFES-20 and ISIC datasets

Abstract: Skin cancer is one of the most common forms of cancer, of which melanoma is the most lethal. Early detection is critical to long term survival rates. The use of machine learning to detect melanoma shows promising results in detecting malignant forms. The use of melanoma detecting models could potentially lead to faster diagnosis, and ensure treatment is applied before the melanoma has spread to other organs. In recent years, many studies have been made on the topic, especially on larger datasets such as ISIC, but there is a lack of studies on how machine learning perform on datasets containing smartphone images of skin lesions. PAD-UFES-20 is a dataset containing 2298 smartphone images of skin lesions, where the malignant forms are all clinically confirmed. The purpose of this thesis was to produce a comparison of the accuracy, precision, recall, and Area under the Curve metrics for several convolutional neural network models that were trained on the PAD-UFES-20 and ISIC datasets to differentiate between melanoma and benign lesions. The goal of this project is to contribute to the knowledge surrounding how neural networks perform in terms of melanoma detection. This is of interest for anyone getting started in the domain of melanoma detection using machine learning, building a melanoma detection application, and for anyone benefiting from earlier diagnosis. A qualitative and inductive approach with the support of quantitative data was adopted. Eight models were selected. They were designed using transfer learning, where the base model acted as a feature extractor, providing input to a dropout layer, and finally to the classification layer, which was retrained for 50 epochs, using the ADAM optimizer with a learning rate of 0.0003. These models were then evaluated against a test set for accuracy, precision, recall and AUC. The results of this study showed that the best performing models in terms of detecting melanoma on the PAD-UFES-20 dataset were Resnet_v2_50 and Mobilenet_v3_Large. Resnet_v2_50 achieved an accuracy of 0.9062, precision of 0.8, recall of 0.6667 and AUC of 0.8846. Mobilenet_v3_Large achieved accuracy of 0.9062, precision of 0.8, recall of 0.6667 and AUC of 0.9295. The best performing model on the ISIC dataset was the Inception_Resnet_v2 model with accuracy of 0.9342, precision of 0.9627, recall of 0.9018 and AUC of 0.9675. The conclusions of this study show that several models achieve promising results on the PAD-UFES-20 dataset. Overall, models that extracted a larger number of features from the base model performed better on the PAD-UFES-20 dataset, whereas this was not the general case on the larger ISIC dataset. It is not too futuristic to claim that machine learning can be applied as complement to standard diagnosis, but the study highlights the need for similar studies on larger and more diverse datasets of smartphone images, from which more transferable results can be obtained.