Convolutional Neural Network Based Skin Cancer Detection (Malignant vs Benign)

Abstract

Skin cancer is very dangerous and deadly diseases in today's world. Between Malignant and Benign skin cancers, Malignant is the deadliest and Benign is curable. Due to the significant growth rate of Malignant and Benign skin cancer, its high treatment costs, and the mortality rate, the need for early detection of skin cancer has been increased. In most cases, these cells are manually identified and it takes time to cure them. In this paper it has been addressed the requirement for a cheap and fast detection of skin disease (Malignant and Benign) applying more effective CNN, PyTorch and to increase the accuracy four different ResNet models has been used. In this method, a pre-trained model named ResNet is used for image classification. It has been used four different version of ResNet model (ResNet18, ResNet50, ResNet101 and ResNet152) to increase the accuracy of our project. ResNet model is a specific type and advance version of deep convolutional neural network. It is better and faster than previously used VGG-16 per-trained model for image classification. Dataset used in this project is collected from Kaggle.com which contains almost 6,599 images to train the model and measure the accuracy. By using different version of ResNet model respectively observed different test result (86.34% for ResNet18 model, 88.78% for ResNet50, 89.09% for ResNet101 and 89.65% for ResNet152). It has been compared the accuracy from our proposed method with the existing method and obtained better accuracy rather than the existing method. The existing system gave an accuracy which is about 83.02% and this system gives more than 89.65% accuracy and it's higher than previously done on skin cancer detection project.