A Pragmatic Note on Evaluating Generative Models with Fréchet Inception Distance for Retinal Image Synthesis

TL;DR

This paper scrutinizes Fréchet Inception Distance (FID) and related feature-distance metrics for evaluating retinal image synthesis, showing that these metrics often fail to predict downstream task performance when synthetic data are used for augmentation in glaucoma classification and retinal layer segmentation. By comparing seven metrics across color fundus photography and OCT with multiple generative models, the authors demonstrate strong internal correlations among metrics but limited or inverted alignment with downstream gains, especially for diffusion models and StyleGAN3. They advocate prioritizing downstream task performance as the primary evaluation criterion and call for developing reliable proxy metrics that correlate better with practical utility and data-centric learning strategies. The findings highlight a significant practical impact: metric-based evaluation alone is insufficient for biomedical data augmentation, underscoring the need for task-aware evaluation pipelines in medical image synthesis.

Abstract

Fréchet Inception Distance (FID), computed with an ImageNet pretrained Inception-v3 network, is widely used as a state-of-the-art evaluation metric for generative models. It assumes that feature vectors from Inception-v3 follow a multivariate Gaussian distribution and calculates the 2-Wasserstein distance based on their means and covariances. While FID effectively measures how closely synthetic data match real data in many image synthesis tasks, the primary goal in biomedical generative models is often to enrich training datasets ideally with corresponding annotations. For this purpose, the gold standard for evaluating generative models is to incorporate synthetic data into downstream task training, such as classification and segmentation, to pragmatically assess its performance. In this paper, we examine cases from retinal imaging modalities, including color fundus photography and optical coherence tomography, where FID and its related metrics misalign with task-specific evaluation goals in classification and segmentation. We highlight the limitations of using various metrics, represented by FID and its variants, as evaluation criteria for these applications and address their potential caveats in broader biomedical imaging modalities and downstream tasks.

Figures (6)

• Figure 1: Overview of the generative model pipeline for data enrichment exemplified with retinal images, including generation, evaluation, and downstream task prediction. The discrepancy between generative evaluation metrics and downstream performance is investigated.
• Figure 2:
• Figure 3:
• Figure 4:
• Figure 5: