Identification of SNP mutations linked to melanoma via a CNN network: application to the FGFR2 gene

One of the deadliest diseases in the world is skin cancer, especially melanoma. Detection and identification of this type of cancer is very difficult due to the strong similarity of its morphological characteristics with the characteristics of other skin cancer types, even with nevi. Early detection of melanoma can help in increasing the cure rate by minimizing diagnostic time and saving efforts. Different systems from deep learning are actually developed to help dermatologists early diagnose melanoma. These systems use several types of data for the classification between melanoma and non-melanoma, such as dermoscopic images and DNA or RNA genetic sequences. This paper proposes a new work that aims at recognizing melanoma lesions from non melanoma ones. The method uses the convolutional neural networks. The data we consider here are the SNP protein sequences of the human FGFR2 gene. We transform, then, these SNP sequences from their textual form into scalogram images to be taken as input to the CNN classifier. The images are encoded in different color spaces with the goal of constructing four databanks. For quantitative evaluation, we use different evaluation rates such as: accuracy, precision, specificity, and sensitivity. The values attained are very satisfactory. These results show the performance of the proposed system and demonstrate the existing relationship between the gene FGFR2 and the possible small protein changes.