Computer vision based skin disease detection using machine learning

Abstract

Skin cancer have been the primary focus of this study, as they are one of the most deadly diseases if not diagnosed and treated early. The study will make it possible for computer science and medical science to work together to save lives. Machine Learning, Deep Learning, and Image Processing have already been used to treat skin conditions. Even though Deep Learning, Machine Learning, and Image Processing have all been used in the past to treat skin diseases, we are trying to improve the accuracy of this work by implementing new Image Processing and Machine Learning models.In this study, pre-trained CNN models and combined pre-trained Unet models are used to identify skin cancer and categorize the kind of cancer using datasets with labeled pictures and classes. Although it may develop in areas of your skin that are are not exposed to sunlight, skin cancer is an abnormal growth of skin cells that most frequently affects skin that has been exposed to the sun. Skin cancer is a kind of malignant melanoma, which is a type of cancer. The three most common kinds of skin cancer are melanoma, basal cell carcinoma, and squamous cell carcinoma. Approximately 301 people die from skin cancer each year. It will be difficult to achieve high accuracy if you rely just on the dataset provided by Kaggle. Recognize that not all datasets are balanced. Therefore, this study focuses on identifying several methods to combine Deep CNN models with U-net to produce deep CNN models that have the highest accuracy on both large and small datasets. These methods mostly depend on supervised learning, which involves the use labels and training data from datasets.This research aims to demonstrate how the best model can be used to accurately diagnose Skin cancer and disorders at an early stage. Here, we have classified a unique dataset using U-net model. We further examined the CNN, U-net, pre-trained models, and other Unet based models for example Attention Unet, ResUnet, R2Unet models accuracy to develop an optimum model that may be further customized to a mobile application for widespread usage. On more than 10000 photos of seven different skin conditions compared to healthy skin, we built the architecture. On the basis of the images of our researched dataset, we carefully compared our data and classified it. Finally, using ResUnet, which is the best model for the task at hand, we accurately identified the Seven illnesses with a 87.21% accuracy rate.