Rhi3DGen: Analyzing Rhinophyma using 3D face models and synthetic data

Abstract

Within the realm of medical diagnosis, deep learning techniques have revolutionized the way diseases are identified and studied. However, a persistent challenge has been data scarcity for many disease categories. One primary reason for this is issues related to patient privacy and copyright constraints on medical datasets. To address this, our research explores the use of synthetic data generation, focusing on Rhinophyma, a subclass of Rosacea. Our novel approach uses 3D parametric modeling to create synthetic images of Rhinophyma, addressing the data scarcity problem. Through this method, we generated 20,000 images representing 2000 distinct anatomical deformations of Rhinophyma. This research not only showcases the potential of using 3D parametric modeling for Rhinophyma but hints at its applicability for other diseases with anatomical abnormalities. With just 30 % of this synthetic dataset, we achieved a remarkable 95 % recall in classifying 220 real-world Rhinophyma images. The performance of our classification model is further validated using GradCAM visualisation. Our findings underscore the potential of such techniques to propel medical research and develop superior deep learning diagnostic models when only limited real-world images are available.