Skin Cancer Classification using Transfer Learning

Today, Cancer is one of the major lethal diseases in the world. Globally out of every three cancers diagnosed, one is identified as skin cancer. Some reports suggest that one out of every five Americans might fall prey to skin cancer in the course of their life. Early detection of the disease plays a pivotal role in the treatment of skin cancer. Though these skin lesions can be seen without the help of any external clinical device, it is a challenging task to distinguish between malignant and benign skin lesions as they are alike in their physical appearances. This leads to an increased number of unnecessary biopsies where in one study it was revealed that nearly 5,00,000 biopsies are done in children every year to diagnose a mere 400 melanomas. To tackle this problem and help dermatologists in the diagnosis process, we developed an enhanced image classification model which can act as a preliminary check before moving to a costlier biopsy. This model can identify 7 distinct types of skin lesions. An analysis has been carried out on the HAM10000 dataset. We used transfer learning utilizing multiple pre-trained models, combined with class-weighted loss and data augmentation techniques for the classification process. Experimental analysis shows that the modified ResNet50 model is capable of identifying skin lesion images into one of the seven classes with categorical accuracy, weighted average precision, and recall of 90 percent, 0.89, 0.90, respectively. Our model can be used as a clinical decision support system to help dermatologists in the diagnosis process.