Related Experiment Video
Updated: May 1, 2026
Quantitative Fundus Autofluorescence for the Evaluation of Retinal Diseases
Published on: March 11, 2016
DiFFA: A Diffusion-Based Model for Synthetic Fundus Fluorescein Angiography to Enhance Detection of Diabetic
Jin Wei1, Yupeng Xu1, Tian Niu1
1Department of Ophthalmology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, National Clinical Research Center for Eye Diseases, Shanghai Key Laboratory of Ocular Fundus Diseases, Shanghai Engineering Center for Visual Science and Photomedicine, Shanghai Engineering Center for Precise Diagnosis and Treatment of Eye Diseases, Shanghai, China.
Purpose:
To develop diffusion-based fundus fluorescein angiography (DiFFA), a noninvasive method for generating high-quality FFA images from color fundus (CF) photographs using diffusion models, addressing the limitations of invasive FFA examinations in detecting diabetic microvascular complications.
Methods:
DiFFA was trained on paired CF and FFA images from 4032 patients across 3 hospitals, with external validation using 2000 CF images from the UK Biobank. Four systematic evaluations were conducted: (1) objective image quality assessment using the mean absolute error, peak signal-to-noise ratio, structural similarity index measure, and Fréchet inception distance metrics, (2) subjective authenticity evaluation by ophthalmologists, (3) diagnostic efficiency assessment, and (4) automated classification accuracy for diabetic retinopathy and diabetic kidney disease using VGG16 with and without synthetic FFA images.
Results:
DiFFA achieved strong objective performance (mean absolute error, 109.25,; peak signal-to-noise ratio, 27.677; Fréchet inception distance, 38.466; structural similarity index measure, 0.694). Ophthalmologists showed limited ability to distinguish synthetic from real FFA images (area under the curve, 0.52 ± 0.06 for junior ophthalmologists and 0.57 ± 0.05 for senior ophthalmologists), with no significant quality score differences (P > 0.05). Combining synthetic FFA with CF images improved automated classification, achieving an area under the curve of 0.933 for diabetic retinopathy and 0.849 for diabetic kidney disease internally, compared with 0.891 and 0.792 using CF photographs alone. External validation confirmed improvements (diabetic retinopathy, 0.858 to 0.879; diabetic kidney disease, 0.731 to 0.814).
Conclusions:
DiFFA successfully generates clinically authentic FFA images from non-invasive CF photographs, enhancing diagnostic accuracy while overcoming the invasiveness limitations of traditional FFA examinations.
Translational Relevance:
Diffusion-based fundus fluorescein angiography enables noninvasive early detection of diabetic microvascular complications, improving patient compliance and diagnostic accessibility while reducing examination risks in diabetes screening and monitoring.
