Detection and classification of dermatoscopic images using segmentation and transfer learning

With the increase in the number of cases every year, skin cancer stays as one of the most common cancers worldwide. Although dermatologists have been aided with modern research study in detection of cancer, proper treatment of cancer has been a quite challenging task due to the visual appearance of cells. In this experiment, we have studied segmentation and classification algorithms for early detection of cancerous cells, as early detection may increase survival rate of the affected person. HAM10000 dataset has been utilized in this study which has 10,015 different images into seven different classes. Three different types of segmentation algorithms have been studied in this experiment, U-Net, ResUNet and DeeplabV3+. Among all these, DeeplabV3+ appears to outperform the rest of the algorithms giving an overall accuracy of 96.21%, precision and recall of 93.26% and 93% respectively. Few pre-trained models namely ResNet50, ResNet152, SqueezeNet1.1 and DenseNet121 have been utilized for classification of skin cancer. As per the results obtained from the experiment, ResNet152 outperforms the rest of the three pre-trained models with an overall training and validation accuracy of 81.75% and 78.51% respectively.