Abstract
One of the worst types of cancer is skin cancer, because it spreads to various body parts if it is not treated at early stages. Melanoma accounts for the massive majority of skin cancer related deaths, it is the most well-known forms of skin disease. Initial detection of skin cancer is of interest for medical diagnosis because of visual matching; image classification plays a key role in achieving an appropriate diagnosis of various lesions. In medical assessments, a computer diagnostic system based on deep learning may offer an automatic way to get over this challenging situation. Convolutional Neural Networks are used to increase the skin lesions classification using dermoscopic pictures without the assistance of humans. This research proposes the novel architecture Derma NET, for melanoma classification. In the suggested model, pre-processing techniques like up-sampling is used for augmentation to address the problem of an unequal sample size. From the database the data is divided into Train and Validation. Derma NET extracts various features from the images during training. Relu is a Non activation function is used in the model to extract the complex features. Learning rate is adjusted by utilizing Adam optimizer. Hyperparameter adjustment is done to improve model performance. HAM10000 is publicly available dataset and it is used to train the model. The performance of the suggested model is assessed using ISIC 2017 dataset. To communicate with proposed model an API (Application Programming Interface) is created that runs on Flask and is easily included into the front end. Through this API, users can query the model and get predictions about patient status with reference to skin cancer type. The experiment's findings demonstrate a substantial improvement in classifying skin cancer especially melanoma with 97.9% accuracy, 87% precision, 93% sensitivity, 98% specificity, 90% f1-score and AUC = 1, showcasing its potential for clinical applications in dermatology.