Experimental Comparative Study on Autoencoder Performance for Aided Melanoma Skin Disease Recognition

Abstract

Melanoma is a dangerous and metastatic cancer that may be fatal and it has a high ability to invade other tissues and organs. Early diagnosis is an important reason to recover from melanoma and reduce mortality. So, automatic skin segmentation is considered an enthusiastic study at present. In this paper, we investigate the applicability of deep learning approaches to the segmentation of skin lesions by evaluating five architectures: Deeplabv3plus, Inception-ResNet-v2-unet, mobilenetv2_unet, Resnet50_unet, vgg19_unet by providing a comparative study of those methods. All methods were trained on the ISIC2017 dataset. The methods were trained on the original dataset, and then the dataset was pre-processed for use in training the five methods. We used quantitative evaluation metrics to evaluate the performance of the methods. The Deeplabv3+ architecture showed significant results compared to the rest of the architecture in F1 as high as 89%, Jaccard as high as 83% and Recall as high as 91%.