Translation of Color Fundus Photography into High-resolution Indocyanine Green Angiography Image using Generative Adversarial Networks for Age-Related Macular Degeneration Screening

Abstract

Purpose : To develop and validate a deep-learning model capable of generating realistic early-phase, mid-phase, and late-phase ICGA images from color fundus photography (CF), and evaluate its performance in the screening of age-related macular degeneration（AMD）.

Methods : We developed a deep-learning model using generative adversarial networks (GANs) . The model was trained using 99,002 CF and ICGA image pairs from early-phase, mid-phase, and late-phase in a tertiary center. The quality of generated ICGA images was evaluated objectively by common image generation metrics, and subjectively by two experienced ophthalmologists on 50 sets from internal and external datasets respectively. Two ophthalmologists graded the generated images on a scale of 1-5, considering the global similarity, the fidelity of anatomical structures, and the depiction of fluorescence-based pathological lesions. Moreover, we validated the practicability of the translated ICGA on an external dataset by calculating the area under the ROC curve (AUC) in classifying AMD.

Results : The multiscale structural similarity scores of the translated ICGA images spanned from 0.68 to 0.74, and the subjective quality scores ranged from 1.46 to 2.74 on a five-point scale (1 refers to the image quality of the real ICGA image). Both ophthalmologists indicated similar quality scores with substantial agreement (kappas ranged from 0.79-0.84). Adding the generated ICGA on top of real CF improved AMD classification in the Labelme dataset, with the AUC increased from 0.93 to 0.97 (P value < 0.001).

Conclusions : To the best of our knowledge, this is the first study to develop a CF-to-ICGA translation model based on GANs. The model demonstrates a high level of authenticity in generating anatomical structures and diverse pathological lesions of the choroid and retina. These synthetic ICGA images may have the potential to achieve data augmentation and alleviate data hungry in future deep-learning model training. Nevertheless, image generation by GANs represents an innovative exploration with yet unclear clinical practicality. Real-world prospective trials are necessary before cross-modality generation model enter the clinical workflow.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Figure 1. Flow chart of the study

 

Figure 2. Examples of real and translated ICGA images