Perceptually indistinguishable retinal OCT synthesis and objective computational framework for detection of synthetic images

Generative artificial intelligence (AI) is increasingly deployed in medical imaging to mitigate data scarcity, enable scalable model development, and simulate disease progression; however, as synthetic data approaches photorealism, it introduces a critical vulnerability: the inability to reliably distinguish synthetic from authentic clinical data. In imaging-centric health systems, undetected synthetic data may erode dataset integrity, obscure provenance, propagate bias, and compromise reproducibility and regulatory auditability. Despite these risks, rigorous evaluation of synthetic image detectability remains limited. To this end, we synthesize high-resolution (512 × 512) retinal optical coherence tomography (OCT) B-scans using four representative generative architectures, a Wasserstein Generative Adversarial Network (WGAN), a Denoising Diffusion Probabilistic Model (DDPM), R3GAN, and a Latent Diffusion model (LDM) trained on 88,500 images from the AIREADI dataset and systematically evaluate their perceptual realism and structural fidelity. In particular, perceptual indistinguishability is assessed through two blinded human reader studies involving four retinal specialists, comprising a naïve and post-familiarization evaluation for WGAN-generated images and an independent evaluation of DDPM-generated images. In parallel, we introduce an objective computational framework predicated on gradient-derived (Sobel) representations, with downstream dimensionality reduction via principal component analysis (PCA) and support vector machine based (SVM) classifier, to probe latent structural divergences within feature space. Across both reader studies, retinal specialists demonstrated limited ability to distinguish synthetic from real OCT images. Classification accuracy remained near chance both before (59.5%) and after familiarization (56.5%) for WGAN-generated images, while DDPM-generated images yielded a similar near-chance accuracy (50.3%), indicating that perceptual indistinguishability was observed for both adversarial and diffusion-based image synthesis. In contrast, the proposed objective computational method achieved 99.93%, 99.63%, 99.91%, and 99.83% classification accuracy for WGAN, DDPM, R3GAN, and LDM-generated images, respectively on independent test datasets, demonstrating near-perfect discrimination between real and synthetic images. These findings delineate a fundamental disjunction between human perceptual judgment and computational detectability of synthetic medical images. Although generative models yield clinically plausible OCT scans, they encode latent structural signatures that remain imperceptible to expert readers yet are robustly recoverable with objective computational analysis.