Transformer Based Melanoma Detection Using Deep Learning

Abstract: Detection and classification of skin cancer using dermoscopic images plays an important role in early diagnosis and proper treatment planning. Examining these skin images manually takes a lot of time and it mainly depends on the experience of the doctor. This drawback increases the importance of automated detection systems. This project presents a deep learning framework with the help of transfer learning techniques to detect skin cancer accurately and in a precise manner. The system learns important features from dermoscopic images and improves the overall performance. To improve the capability of the system, a Vision Transformer model is used which analyzes the complete image instead of small regions. The model is able to learn features such as color variation, irregular shapes, uneven texture, and unclear boundaries of skin lesions. Before giving the images to the model, they are prepared in a simple way so that the system can learn better. First, the skin images are made ready so that the model can understand them properly. We resize some images and add a few more so the system can understand the data properly. The model was developed using Python with PyTorch. After the training part is finished, the system is given a skin image and it tries to identify what type it is. It checks the image and gives the result as either normal or cancer. When we tested the system, it worked fine in most situations and did not take much time to give the output.

Keywords: Dermoscopy Images, Skin Cancer, Melanoma Detection, Transfer Learning Model, Vision Transformer, Pytorch.