Repurposing the Image Generative Potential: Exploiting GANs to Grade Diabetic Retinopathy

Isabella Poles, Eleonora D'arnese, Luca G. Cellamare, Marco D. Santambrogio, Darvin Yi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 2305-2314

Diabetic Retinopathy (DR) is a common cause of irreversible vision loss in the working-age population. Automatic DR grading allows ophthalmologists to provide timely treatment to numerous patients. However developing a robust grading model needs large balanced and annotated data which poses challenges in the collection. Moreover data augmentation often fails to generate diverse data necessitating alternative approaches such as Generative Adversarial Networks (GANs). However GANs often operate with low-resolution images as a result of their costly training process. Therefore we present a novel method that repurposes the discriminator of an unconditional Progressive GAN leveraging the generative knowledge gained for DR grading. Furthermore a new Log-Likelihood Inception Distance (LLID) metric estimates the similarity between one synthesized and a set of real images thereby capturing human judgment more effectively. Our method is validated through extensive experiments on three public datasets outperforming the baseline classifiers' performance by 12.5% and 14.33% average accuracy on small data regimes and when combined with state-of-the-art methods on large datasets respectively. Additionally LLID reproduces the comprehension ability of most of our Visual Turing Test participants enabling differentiation between a synthesized image and a set of reference images with 82.88% accuracy. This confirms the quality of generated images and the metric consistency with human decision-making mechanisms.