CNN and Metadata for Classification of Benign and Malignant Melanomas

Abstract

Skin cancer is detected in skin lesions. The most common skin cancer is melanoma. Skin cancer is increasing in several parts of the world. Due to the above, it is important to work on the classification of melanomas, in order to support the possible detection of malignant melanomas that cause skin cancer. We use Convolutional Neural Networks (CNN) for the classification of melanomas. We use images available from International Skin Imaging Collaboration (ISIC). We created a repository of 1000 images and did training with a sequential CNN to obtain two categories: benign and malignant melanomas. In the first instance we obtained results of 94.89% accuracy and 82.25% in validation. In the second instance we created another repository of 600 images for the method that we propose that consists in adding metadata within the same pixel matrix of the image in each RGB layer. The image was shown with a band of colors at the bottom. We made training with the CNN using images with metadata and achieved the results: 98.39% of accuracy and 79% of validation. Therefore, we conclude that adding the metadata repeatedly to the pixel matrix of the image improves the results of the classification.