Skin Cancer LesionClassification Using Transfer Learning based Fine Tuned Deep Neural Networks

Abstract

Skin cancer is a foremost concern for human society, with pigments producing skin color turning carcinogenic and causing the disease to develop. Early detection of lesions is crucial to cure skin cancer, but diagnosis can be challenging as many pigments can appear similar. To assist dermatologists in this process, Artificial intelligence-based automated tool development has advanced significantly.The aim of this research is to shorten the time it takes to get a skin biopsy report from a days or more to just a couple of hours. This has the potential to positively impact millions of people. To make the tool accessible to everyone, an easy-to-use website is being developed. Users or dermatologists can upload patient demographic information along with the skin lesion image. Further,DenseNet201 or InceptionV3model analyses the data and returns prompt results. Additionally, a basic infographic page is being developed to provide a generalized overview of melanoma and steps to use the online tool to get results. By fine-tuning all the layers in the InceptionV3 and DenseNet201 the results achieved are very satisfactory. InceptionV3 model after fine-tuning all the layers the validation accuracy achieved is 86.91% and testing accuracy is 86.821% and the best model found is by fine-tuning all layers of DenseNet201 model where the validation accuracy achieved is 86.697% and testing accuracy is 87.724%.