Generating Synthesized Fluorescein Angiography Images From Color Fundus Images by Generative Adversarial Networks for Macular Edema Assessment

Abstract

Purpose: To assess the feasibility of generating synthetic fluorescein angiography (FA) images from color fundus (CF) images using pixel-to-pixel generative adversarial network (pix2pixGANs) for clinical applications. Research questions addressed image realism to retinal specialists and utility for assessing macular edema (ME) in Retinal Vein Occlusion (RVO) eyes.

Methods: We used a registration-guided pix2pixGANs method trained on the CF-FA dataset from Kham Eye Centre, Kandze Prefecture People's Hospital. A visual Turing test confirmed the realism of synthetic images without novel artifacts. We then assessed the synthetic FA images for assessing ME. Finally, we quantitatively evaluated the synthetic images using Fréchet Inception distance (FID) and structural similarity measures (SSIM).

Results: The raw development dataset had 881 image pairs from 349 subjects. Our approach is capable of generating realistic FA images because small vessels are clearly visible and sharp within one optic disc diameter around the macula. Two retinal specialists agreed that more than 85% of synthetic FA images have good or excellent image quality. For ME detection, accuracy was similar for real and synthetic images. FID demonstrated a 38.9% improvement over the previous state-of-the-art (SOTA), and SSIM reached 0.78 compared to the previous SOTA's 0.67.

Conclusions: We developed a pix2pixGANs model translating FA images from label-free CF images, yielding reliable synthetic FA images. This suggests potential for noninvasive evaluation of ME in RVO eyes using pix2pix GANs techniques.

Translational Relevance: Pix2pixGANs techniques have the potential to assist in the noninvasive clinical assessment of ME in RVO eyes.

Introduction

Retinal vein occlusion (RVO) is one of the leading causes of visual impairment.^1–3 The estimated prevalence of RVO ranges from 0.26% to 3.39%, and it is associated with factors of older age and other systemic diseases.⁴ We recently reported that RVO is a common cause of blindness secondary to cataract and age-related macular degeneration (AMD) in a Tibetan population.⁵ In RVO, the fundus may show retinal hemorrhages, dilated tortuous retinal veins, cotton-wool spots, optic edema, and macular edema (ME). Among these, ME is the most important cause of visual impairment in RVO.⁶ As new advanced treatment, such as intravitreous injections of antivascular endothelial growth factor (VEGF) agents,⁷ have been shown to significantly improve visual acuity in eyes with RVO-associated ME, it is important to identify vision-threatening complications, especially ME, early. 

Color fundus (CF) and fluorescein angiography (FA) imaging have been the gold standard in RVO diagnosis for decades.⁸ CF imaging is the most popular technique in ophthalmology practice to document retinal disorders because of its easy accessibility and noninvasiveness. The main limitation of CF is that it does not visualize microstructural changes in the retina. FA allows the direct imaging of blood vessels and blood flow in the retina. Furthermore, FA provides very useful information on the vascular status of the retina, presence of neovascularization, delayed venous filling, retinal leakage, and ME, which make it the most widely used test to establish the diagnosis of RVO.⁹ Unfortunately, it is also well established that FA has a high rate of adverse reactions at 10%. One in 200,000 FA may result in a patient's death.¹⁰ 

Generative models¹¹ based on deep learning (DL) methods, such as generative adversarial networks (GANs), have gained much attention in medical imaging analysis. Several studies have already shown promising results of GANs synthetic images in a wide range of applications, such as synthesizing radiological images of the spine, generating magnetic resonance images from brain computed tomography images.^12,13 In ophthalmology, we have shown that GANs synthetic ocular coherence tomography (OCT) images have good quality for clinical evaluation and can also be used for developing deep learning algorithms.^14–16 Furthermore, we demonstrated that cross-modality image transfer between anterior segment optical coherent tomography (ASOCT) and ultrasound biomicroscopy is possible by using cycle-consistent GANs (CycleGANs).¹⁷ However, cross-modality image synthesizing between CF and FA images is a challenging problem because the different imaging mechanisms cause high variability of vessels or retinal microstructure appearance. To address this issue, we recently proposed a cross-modal image registration approach,¹⁸ and our preliminary results are promising in predicting FA from CF images. In this study, we aim to build a GANs-based deep learning model for the domain transfer from CF to FA images. We then conducted experiments to evaluate the following research questions: whether the image quality of synthetic FA images appears realistic to retinal specialists and whether synthetic FA images can be used for ME assessment in a clinical dataset. 

Methods

Image Datasets

This study was approved by the Institutional Review Board of Kandze Prefecture People's Hospital (identifier, XHEC-D-2021-114) and conducted in accordance with the principles of the Declaration of Helsinki as amended in 2008. Written informed consent was obtained from all participants. 

We retrospectively collected the CF and corresponding FA images from eyes with normal and RVO from Kham Eye Centre, Kandze Prefecture People's Hospital. The raw development dataset consisted a total of 881 pairs of images from 349 subjects. All subjects were diagnosed as RVO with/without macular edema between September 16, 2018, and April 10, 2022. All subjects underwent a full ophthalmic examination, including best-corrected visual acuity, refraction, slit-lamp examination, intraocular pressure, and fundus examination by a fellowship-trained retinal specialist. We excluded patients with a prior history of trauma, intraocular tumor, or previous retinal surgery. Eyes with dense cataract and diabetic retinopathy were also excluded. The fellow eyes of RVO eyes were considered as non-RVO control if they fulfill the following criteria: best-corrected visual acuity ≥ 20/40; a presenting intraocular pressure < 21 mm Hg on non-contact tonometry; no previous history of trauma, intraocular tumor, retinal surgery, or wet AMD and proliferative diabetic retinopathy. We used a Zeiss non-mydriatic fundus camera (VISUCAM; Carl Zeiss Meditec, Jena, Germany) to capture both CF and FA images. This device's capabilities have been detailed in prior studies.¹⁹ Essentially, the camera's sensitive sensors enable the use of attenuated flash settings, which reduce sustained pupillary constriction and facilitate the capture of high-resolution two-dimensional ocular images. All CF images were of the posterior pole and obtained at 45°. For FA images, we only selected images with late phase, which better showed the extent and severity of ME. All images were subjected to a grading system consisting of two layers of trained graders of increasing expertise (Fig. 1). The first layer of graders consisted of three trained medical students and nonmedical undergraduates. They conducted initial quality control according to the following criteria: (1) inadequate field definition was defined as either a small-pupil artifact is present, or at least one of the macula, optic disc, superior temporal arcade, or incomplete inferior temporal arcade; (2) inadequate image clarity was defined as small vessels that are not clearly visible within one optic disc diameter around the macula or 50% or more of the area was obscured.²⁰ The second layer of graders consisted of three Chinese board-certified ophthalmologists (with more than five years of experience in the retinal subspecialty). Each grader independently reviewed the medical record and graded each pair of CF and FA images into normal, RVO, and RVO with ME. ME caused by RVO was diagnosed by fundus examination and FA workups. The fundus showed an increase in macular thickness, fluid, or exudates. FA can show vascular leakage in the early or mid-phase and filling of cystic spaces in the late phase.²¹ In the event of discordant annotations between the two grader layers, the final classification for an image was determined through a subsequent review by an independent senior board-certified ophthalmologist (more than 20 years of experience in the retinal subspecialty). This arbitrating clinician's evaluation was deemed conclusive, ensuring that each image was assigned to an appropriate category with a high degree of confidence. 

Figure 1.

 

Standards for Reporting Diagnostic Accuracy Studies diagram showing dataset establishment, DL models development, and validation.

Image Registration and Preprocessing

The initial phase of our research involved meticulous image registration and preprocessing to ensure precise alignment of CF and FA imagery, thereby eliminating spatial incongruities. To address common acquisition challenges such as non-uniform illumination and diminished contrast, we judiciously applied data augmentation strategies. Histogram equalization was used to enhance image contrast, whereas the removal of the red channel from fundus photographs effectively suppressed interference from choroidal vasculature, thereby streamlining the delineation of vascular features. In the ensuing registration process, we extracted feature points from corresponding fundus images, leveraging gradient information from surrounding regions to ascertain the principal orientation of these points and construct their descriptor vectors for nuanced characterization.²² Using the Best-Bin-First matching algorithm, we identified analogous feature points across image pairs, thus establishing a mapping relationship that facilitated spatial transformation and successful registration of FA images with their fundus counterparts.²³ 

Pix2pix Generative Adversarial Network (Pix2pixGANs) Architecture and Generating Synthesized FFA Images

We recently reported that a DL algorithm (pix2pix generative adversarial network) can predict FA from CF images.¹⁸ Our discourse encompasses the foundational principles of GANs, their convergence properties, and their applicability to tasks involving the prediction of FA from CF images. Adversarial training enables the generator to iteratively refine its ability to accurately map CF photographs to their corresponding FA representations. 

The GANs framework used in this study is depicted in Figure 2. We used the pix2pixGAN architecture to facilitate learning the transformation from CF to FA images. As previously delineated by Isola et al., the pix2pixGAN framework possesses two salient characteristics: first, it incorporates a U-Net-based generator that captures and preserves both low-level and high-level features from the input image to the output²⁴; second, it employs a PatchGAN discriminator, which scrutinizes patches within the output image rather than the entirety, thereby encouraging the generation of detailed and crisp images.²⁵ 

Figure 2.

 

Framework of pix2pix generative adversarial network.

The primary challenge in modality conversion is the translation of high-resolution inputs into corresponding high-resolution outputs. Notwithstanding disparities in appearance, these modalities often share common underlying distributions and are anchored by analogous anatomical structures. Conventional approaches have predominantly relied on encoder-decoder networks, which compress inputs to a bottleneck layer through consecutive neural network layers. However, given the substantial overlap of information between the input and output in numerous modality conversion tasks, it is imperative to facilitate direct data flow across the network. For example, salient edge positions in color fundus images typically remain invariant between modalities. As deep neural networks become increasingly complex, they encounter performance degradation—a phenomenon known as the degradation problem, which stems from overfitting or excessive model depth. This degradation intensifies with the accumulation of nonlinear transformations across layers, resulting in the attenuation of original feature information. To counteract this, our predictive model incorporates a GAN generator with a U-Net architecture augmented by skip connections, as illustrated in Figure 2. 

The discriminator component of our GAN, known as PatchGAN, is designed to prioritize high-frequency details within the image. Traditional L1 and L2 loss functions excel at capturing low-frequency components but falter with respect to high-frequency nuances. The PatchGAN architecture addresses this by focusing on localized patches of the image during training. This approach models the image as a Markov random field, with the discriminator's output derived from averaging the responses across all N × N blocks. This method assumes independence between pixels separated by distances exceeding the diameter of an image patch—a principle commonly applied in texture and style modeling, as explored by Li et al.¹³ In contrast to conventional discriminators that render a single evaluative score for an entire image, our PatchGAN's N × N matrix-based assessment emphasizes finer details within the image, thereby showcasing the structural superiority of this methodology. 

We followed the established protocol described in reference, opting for alternating updates of the discriminator (D) and generator (G), as opposed to training G to minimize, as per original GAN literature recommendations.²⁶ Instead, we trained G to maximize. To moderate the learning rate of D during optimization, we reduced the objective function by half, thereby decelerating D's acquisition of knowledge. We used minibatch Stochastic Gradient Descent in conjunction with the Adam optimizer, setting the learning rate at 0.0002 and defining the momentum terms as β1 = 0.5 and β2 = 0.999. These strategic modifications were instrumental in stabilizing the training process, thereby enhancing model performance. The network was subjected to an extensive training regimen on an NVIDIA HGX A100 4-GPU configuration, spanning over 200 epochs, with each epoch consisting of 600 iterations. This comprehensive training protocol ensured that our model effectively internalized the complex mapping between CF and FA images, successfully navigating the challenges associated with the high variability observed in ocular conditions and image quality. 

Evaluation of Pix2pixGANs Synthetic FA Images for Macular Edema Assessment

We first performed a visual Turing test²⁷ to assess whether the image quality of synthetic images appeared realistic to retinal specialists and did not introduce novel artifacts (Fig. 3). We used a laptop to display the pix2pixGANs synthetic FA images (256 × 256 pixels) from the testing dataset. We modified an image quality grading scheme from Tavakkoli's³⁰ study for evaluating the GANs synthetic FA images (Table 1), and these sample figures illustrate the varying quality of images generated by our model (Fig. 4). Two retinal specialists (each with more than five years of retinal experience, respectively) from the same center manually graded both synthetic and real FA images. To avoid confirmation bias, we did not tell two retinal specialists that there were any synthetic images in the FA images. 

Figure 3.

 

Evaluating the realism of synthetic FA images: A visual Turing test.

Table 1.

 

Synthetic and Real FA Image Quality Grading Scheme

 

Figure 4.

 

Examples of synthetic FA images with different quality levels.

We then assessed whether synthetic FA images could be used for assessing ME in RVO eyes. Both retinal specialists performed manual annotation of 76 images, which included equal numbers of real and synthetic RVO images with or without ME. For each FA image, they answered the following question: “For each FA image of RVO with ME, is there evidence for ME in corresponding synthetic FA image as well?” 

After completing the grading, we informed the retinal specialists that the FA dataset was composed of a mixture of real and synthetic images. We then asked two retinal specialists to discriminate between real and synthetic FA images. 

Finally, we chose the Fréchet inception distance (FID) and structural similarity measures (SSIM) to evaluate the synthetic images. The FID²⁸ is a popular metric used to evaluate the quality and diversity of generated images in the field of generative artificial intelligence. It calculates the Wasserstein-2 distance in the feature space of deep neural network, which provides a quantitative measure of the similarity between the distribution of real images and the distribution of generated images (a smaller FID value refers to better quality). The SSIM²⁹ is a metric used to measure the similarity between two images. It assesses the perceived quality of an image by considering three key components: luminance, contrast, and structure (a SSIM value close to 1 refers to better similarity). FID and SSIM complement each other in the evaluation of generated image quality: the former focuses more on semantic-level information and considers both generation quality and diversity, whereas the latter measures the proximity of generated images to the ground truth in multiple dimensions of morphological features. 

To further enhance the robustness of our evaluation, we also incorporated the Kernel inception distance (KID). KID is a method that measures the statistical differences between generated and real images by comparing their distributions in the feature space, thereby evaluating the performance of generative models. Specifically, we compared the input fundus images and the generated FA images, the input fundus images and the real FA images, and the generated FA images and the real FA images. 

Statistical Analysis

We used metrics including accuracy, sensitivity, specificity, negative predictive value, positive predictive value, and κ score for the Turing test. The margin of error, when added to the value, provides the 95% confidence interval. All statistical tests were performed using the using Python package, version 3.9 (Python Software Foundation). 

Results

After image grading and preprocessing, the training and testing dataset included 572 and 143 pairs of CF and FA images respectively. Figures 5A and 5B show the example of real FA images and corresponding synthetic FA images. Overall, our approach is capable of generating realistic FA images, because small vessels are clearly visible and sharp within one optic disc diameter around the macula. 

Figure 5.

 

Examples of color fundus image, FA, and corresponding synthetic FA.

Table 2 demonstrated that synthetic images have approximately the same fraction of different levels of images' quality as real images graded by two retinal specialists (all P > 0.05, Pearson χ²). Both retinal specialists agreed that more than 85% of synthetic FA images have good or excellent image quality (90.2% vs. 91.6%). 

Table 2.

 

Synthetic and Real FA Images’ Quality Grading by Two Retinal Specialists

 

Table 3 presented the diagnostic performance metrics of the retinal specialists on real versus synthetic FA images. To detect ME, two retinal specialists had accuracy that was similar on real versus synthetic FA images (83.9% vs. 86.7 % for retinal specialist 1 and (85.3% vs. 81.2 % for retinal specialist 2, respectively). 

Table 3.

 

Retinal Specialist Grading Macular Edema Versus Non-Macular Edema in Synthetic and Real FA Images With Retinal Vein Occlusion

 

Table 4 summarized the Turing test of two retinal specialists to distinguish real versus synthetic FA images. Our results showed that they had limited ability to discern real from synthetic FA images, and the accuracy for distinguishing between true and fake was 56.3% and 54.4%, respectively. 

Table 4.

 

Synthetic Versus Real FA Images Distinguished by Two Retinal Specialists

 

We used FID and SSIM as metrics for quantitative evaluation and compared them with those in the previous works (Table 5).^30,31 As shown in Table 5, our method achieved superior image synthesis quality compared to previous approaches in both the FID and SSIM. For the FID, our method obtained an improvement of 38.9% than the previous state of the art. For the SSIM, our method obtained a significantly higher value (0.78) against the previous state of the art (0.67). 

Table 5.

 

Comparing FID and SSIM With Those in the Previous Works

 

Furthermore, we computed the KID between various groups of images. Specifically, we found that the KID between the input fundus images and the generated FA images was 69.99, whereas the KID between the input fundus images and real FA images was 420.18. Additionally, the KID between the generated FA images and real FA images was 434.05. These KID values are relatively high, indicating significant differences between the compared groups. However, this does not necessarily imply poor quality of the generated images. The high KID values are likely due to the different modalities of the compared images (fundus vs. FA), which inherently have different characteristics and features. 

In the absence of KID values from other comparative works, we conducted a self-validation by comparing the KID scores between synthetic FA images and real FA images, and between the input fundus images and real FA images. The similarity rate between these KID values is 96.8%, calculated based on the values 420.18 and 434.05. This high similarity rate suggests that our synthetic FA images closely resemble the real FA images. 

We also noticed several failure cases as shown in Figure 5C. For example, if an input CF image has blurred fundus because of either cataract or vitreous opacity, the proposed pix2pixGANs would not be able to generate detailed microstructure of the macula. 

Discussion

In the current study, we presented pix2pixGANs-based DL techniques for directly generating FA from CF images. Our study demonstrates that pix2pixGANs are capable of synthesizing realistic FA images. These synthetic images can help in identifying macrostructural changes indicative of conditions such as ME in RVO eyes, although the resolution of pathophysiological characters of the retina may not fully match those obtained with real FA. The pathophysiological characters of retinal obtained with real and synthetic FA images cannot be considered interchangeable because there is fair to excellent correlation between them. The potential application of this technique is promising because it might alleviate issues of adverse reactions during FA, especially in remote areas with poor medical accessibility. 

A generative model, such as GANs or diffusion model, is a type of machine learning model that can learn to generate realistic data from a given distribution, such as images, text, or audio. Generative models have been widely adopted in medicine for various applications, such as synthesizing medical images, augmenting data sets, detecting anomalies, and generating diagnoses. For example, several other and our studies have demonstrated the capability of GANs to generate synthetic CF of AMD, ASOCT of angle-closure, and OCT of retinal diseases. Generative models can also help overcome the challenges of data scarcity, privacy, and complexity in the medical domain. Our previous studies showed that, even in a small labeled dataset, semi-supervised GANs could detect the eye disorders by different imaging modalities, and the performance of the semi-supervised GANs was better than (or at least equal to) that of a supervised DL model.^16,32 More recently, we present that cross-modality image transfer between ASOCT and ultrasound biomicroscopy using CycleGANs. CycleGANs³³ is a type of unsupervised (unpaired collections of images) machine learning technique used for mapping different image domains. Different from CycleGANs, pix2pixGANs use conditional GANs and learn a mapping from an input image to an output image. Training pix2pixGANs is more challenging because the networks need tightly-correlated images (such as pair CF and corresponding FA images in our study). Some researchers showed that pix2pix has better performance than CycleGANs under the same generators.³⁴ More specifically, we aligned the original size CF images of the same person with the FA contrast image, and then used an image mask that preserved only the crossover area between them. Our method produces consistently lower FID measures when producing FA images compared to previous studies.³⁰ SSIM is considered to be correlated with the quality perception of the human visual system. Our proposed method achieved higher SSIM measures comparing to those of Wang et al.³¹ and Isola et al.²⁶ 

Image registration is a critical preliminary step in our methodology because it involves transforming various datasets into a unified coordinate system. This step is pivotal in ophthalmology because it ensures the spatial congruence necessary for accurate diagnosis and treatment planning, particularly for retinal pathologies such as RVO. For instance, in the context of RVO, ophthalmologists identify static retinal pathologies such as macular edema from CF. FA imaging provides crucial information regarding vascular leakage, which is pivotal for guiding treatments such as laser therapy. Alongside this, during model training, aligning CF with FA images allows the model to minimize spatial discrepancies and facilitate accurate modality transformation. To mitigate the risk of bias, we have implemented the following measures: First, we have adopted automated registration algorithms that rely on objective anatomical landmarks, thereby reducing the subjectivity associated with manual registration techniques. Second, we have compiled a diverse and representative dataset that encompasses a wide range of anatomical variations and pathological conditions, which enables our model to generalize across different scenarios rather than conforming to specific registration characteristics. Third, we have conducted robust training of our generative model on a comprehensive dataset that includes a variety of eye conditions, which facilitates the learning of FA image generation based on the inherent characteristics of CFs. Finally, we have engaged in continuous evaluation, using an independent test set to monitor and rectify any model propensity for registration-induced biases. Post-generation, we have also performed a meticulous evaluation of the synthesized FA images to ensure that they consistently and accurately reflect the pathology present in the CFs, without being influenced by the registration process. These measures are taken to ensure that our registration-guided GAN framework leverages the benefits of image registration while minimizing the potential for bias introduction. 

Both CF and FA are commonly performed and extraordinarily valuable diagnostic procedures in ophthalmology. The reported frequency of all adverse reactions of FA varies from 0.6% to 16.1%.³⁵ These reactions included nausea, vomiting, hives, and other reactions (dyspnea, syncope, excessive sneezing). More severe adverse reactions involved anaphylaxis, myocardial infarction, pulmonary edema, or seizures. Therefore recent reports suggested that other noninvasive imaging modalities, such as swept-source widefield OCT angiography, could be an alternative tool to evaluate pathological retinal and choroidal vascular changes compared to FA.³⁶ However, the use of OCT angiography is expensive and not always available at clinical centers (the current study was performed in a remote area of China). Noninvasive and feasible techniques with both high quality and diagnostic capability would be valuable to improve the ability of diagnosing and treating RVO without performing FA. To address this need, we leveraged a pix2pixGANs technique to generate the high-quality FA images that appeared to be realistic to retinal specialists. In the Turing test, two retinal specialists can only achieve nearly 50% accuracy when grading real or synthetic FA images, indicating that the retinal specialists were no better than random chance at correctly identifying synthetic FA images. 

According to the retinal vein occlusions preferred practice pattern 2020 by the American Academy of Ophthalmology,³⁷ comprehensive ocular examination and retinal imaging should do the following: (1) distinguish RVO as either BRVO or CRVO, (2) evaluate for ME, (3) estimate the degree of retinal ischemia, and (4) evaluate for retinal or iris neovascularization. We conducted two experiments to evaluate whether synthetic FA images can be used for ME assessment in a clinical dataset. The diagnostic performance of the FSL shows comparable results in AUC and accuracy for meta-training using only real FA images or synthetic images. Our results were encouraging in that the DL system still achieved respectable results with performance not far from human performance based on earlier studies even though machine training was based completely on synthetic data. However, the validation of synthetic FA images extends beyond technical feasibility; it necessitates a comprehensive comparison against the clinical efficacy of these images, including their direct evaluation against raw CF images, generated FA images, or real FA images. Furthermore, the potential impact of this validation must be considered within the broader context of patient safety in ophthalmology. Conventional FA carries a risk of adverse reactions, with mild to severe reactions occurring in up to 10% of cases and a rare but severe risk of fatality in approximately one in 200,000 cases. By offering a noninvasive alternative for generating FA-equivalent images, our work aims to maintain the diagnostic benefits of FA while significantly enhancing patient safety. 

While our study illustrates the potential of utilizing GANs for synthesizing FA images from CF photographs to assess conditions such as ME, it also encounters several limitations that warrant further discussion. The proposed pix2pixGANs model can only generate FA images with 256 * 256 pixels on the posterior pole of retina. Some small lesions, like dot hemorrhages or collateral vessels, may need higher resolution to evaluate. Previous studies also demonstrated the size of the non-perfusion area in the peripheral retina beyond the area reached by 55° FA that is predictive of the risk of developing neovascularization remains unclear.³⁸ Our previous study reported that it is possible to generate medical images with higher resolutions (e.g., 1024 × 1024 or above for OCT). Our further study will involve generating small lesions using ultrawide-field FA images. Second, the whole FA procedure included several different phases (e.g., choroidal, arterial, capillary, venous, and late phases). The aim of this preliminary study was to explore the capability of GANs for the domain transfer from CF to FA. We will present the GANs synthesizing FA images with different phases in our future study. Another limitation is the homogeneity of our dataset, which was sourced from a single center, employing consistent methodologies and equipment. This uniformity may lead to an overestimation of the model's performance due to potential domain shift vulnerabilities. Acquiring sufficient FA images from eyes without pathology for a true control group is inherently challenging; consequently, we utilized the contralateral eyes of patients with RVO as a practical substitute. Future studies should endeavor to validate the robustness of pix2pixGANs algorithms across diverse datasets, including those from multiple medical centers and encompassing a wider array of ocular conditions. As our pix2pixGANs have been developed and deployed on laptops, this feature ensures their utility in remote areas, facilitating their real-world adoption. Moreover, the efficacy of our model is dependent on the quality of input CF images, and its generalizability across diverse retinal pathologies remains to be explored. To overcome these limitations and further enhance image quality, we propose exploring attention mechanisms that mimic radiologists’ diagnostic processes. Specifically, integrating techniques such as gradient-weighted class activation mapping could significantly improve the detail and accuracy of predicted images, particularly in areas of pathological interest. Future research will also consider anomaly detection to refine the generation of lesion-specific details, pushing the boundaries of what our deep learning model can achieve in the domain transfer from CF to FA images. 

Conclusions

We proposed and validated pix2pixGANs for synthesizing FA images using CF images as the inputs. Our results showed that synthetic FA images appear realistic to retinal specialists. Furthermore, synthetic FA images showed promise in clinical evaluation, such as ME assessment. We expect the findings of this study to be used as an auxiliary technique for clinical evaluation of retinal disease, further improving patient safety by reducing the need of FA. 

Acknowledgments

Supported by the National Natural Science Foundation of China (81371010), Hospital Funded Clinical Research, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (21XJMR02) and Hospital Management Research Program of Institute of Hospital Development Strategy, China Hospital Development Institute, Shanghai Jiao Tong University (HDSI-2022-A-001). 

Disclosure: X. Xie, None; D. Jiachu, None; C. Liu, None; M. Xie, None; J. Guo, None; K. Cai, None; X. Li, None; W. Mi, None; H. Ye, None; L. Luo, None; J. Yang, None; M. Zhang, None; C. Zheng, None