Skin Cancer Classification from Dermatoscopy Images Using Deep Neural Network

Abstract

Skin cancer is one of Indonesia's most common malignant cancers and can cause death. A dermatologist uses a sample and microscope procedure to diagnose skin cancer manually. Nevertheless, this procedure is laborious, and there is a chance that the biopsy will be done incorrectly. More than 90% of cases can be cured with an early diagnosis. However, less than 50% can be fixed with a late diagnosis. This study suggests a convolutional neural network (CNN) approach to classify skin cancer using an Alexnet architecture. A dataset called HAM10000 ("Human Against Machine with 10000 training images") was taken from the International Skin Imaging Collaboration (ISIC) dataset and used in the experiment. We employed 9000 data points in this investigation that were suitable for identification after preprocessing. With a percentage distribution of 80% training data and 20% validation data, the dataset will be used for both training and validation. Thus, 7200 photos of skin cancer were used as training data. Simultaneously, 1800 photos are used as validation data. The test findings demonstrate that the system has an accuracy rate of 80% and a loss value of 0.5459 in classifying skin cancer based on its type. According to system performance data, the generated model may substitute early skin cancer detection.