DermoCC-GAN dataset

Published: 26 July 2022| Version 1 | DOI: 10.17632/6nr7symnjj.1
Massimo Salvi, Francesco Branciforti, Federica Veronese, Elisa Zavattaro, Vanessa Tarantino, Paola Savoia, Kristen Meiburger


This repository contains the image dataset and the scripts used to develop the DermoCC-GAN for standardizing dermatological images: - Salvi M., Branciforti F., Veronese F., Zavattaro E., Tarantino V., Savoia P., and Meiburger K. M. , "DermoCC-GAN: A new approach for standardizing dermatological images using generative adversarial networks", Computer Methods and Programs in Biomedicine 2022 Abstract: Dermatological images are typically diagnosed based on visual analysis of the skin lesion acquired using a dermoscope. However, the final quality of the acquired image is highly dependent on the illumination conditions during the acquisition phase. This variability in the light source can affect the dermatologist's diagnosis and decrease the accuracy of computer-aided diagnosis systems. Color constancy algorithms have proven to be a powerful tool to address this issue by allowing the standardization of the image illumination source, but the most commonly used algorithms still present some inherent limitations due to assumptions made on the original image. In this work, we propose a novel Dermatological Color Constancy Generative Adversarial Network (DermoCC-GAN) algorithm to overcome the current limitations by formulating the color constancy task as an image-to-image translation problem. By training the generative adversarial network with a custom heuristic algorithm that performs well on the training set, the model learns the domain transfer task (from original to standardized image) and is then able to accurately apply the color constancy on test images characterized by different illumination conditions. The proposed algorithm outperforms state-of-the-art color constancy algorithms for dermatological images in terms of normalized median intensity and when using the color-normalized images in a deep learning framework for lesion classification (accuracy of the seven-class classifier: 79.2%) and segmentation (dice score: 90.9%). In addition, we validated the proposed approach on two different external datasets with highly satisfactory results. The novel strategy presented here shows how it is possible to generalize a heuristic method for color constancy for dermatological image analysis by training a GAN. The overall approach presented here can be easily extended to numerous other applications.



Dermatology, Image Normalization, Deep Learning, Image Analysis