A Deep Learning Model for Skin Lesion Analysis Using Gaussian Adversarial Networks

Abstract

Computer assisted radiology becomes an interdisciplinary domain between mathematics, medicine and engineering. Tumor detection, analysis, classification are main problems in digital radiology for diagnosis and follow-up. A physician or an oncologist involves in the care of patients by regarding detailed reports of carcinoma in situ that analyze the pathology of suspicious lesions. Deep learning applied to several fields in medicine is considered as an intervention for oncology. Even if the final treatment of the lesion is decided by the oncologists or the surgeons in a case of resection, image based analysis of lesions (benign or malign) promises automated decision making for radiology. Skin lesion detection and classification are current challenges in medical image analysis. Dermatologic image processing benefits from the evaluation scores of neural nets. Gaussian Adversarial Networks (GAN) bring a new architecture in machine learning by adding generator and discriminator steps in data analysis. In this article, GAN architecture has been implemented on two dimensional skin lesion images. After the preprocessing, colored images have been trained in GAN. The experiment setup has been enriched by adding incremental noise on tumor images before GAN training. The evaluation has been tested through accuracy, sensitivity, specificity, Dice coefficient and Jaccard coefficient parameters. In conclusion, test results showed that GAN architecture provides a robust approach in skin lesion analysis.