Generative adversarial networks (GAN) for pre-dilation retinal photograph quality enhancement

Retinal image quality significantly affects the performance of diagnostic artificial intelligence (AI) models and is typically improved with pupil dilation in clinical settings. However, in real-world settings where dilation is not feasible, suboptimal image quality remains a challenge for AI deployment. In this study, we fine-tuned a CofeNet model to enhance the quality of undilated retinal images.

We performed pixel-level alignment and fine-tuned a generative CofeNet model using 313 paired and spatially aligned undilated and dilated retinal images. Model performance was evaluated on an internal test (120 image pairs) and two external tests. Image similarity was assessed using peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Image gradability was evaluated for undilated and model-enhanced images by three graders on a three-point scale (gradable, force gradable, ungradable), with consensus-based grading. Inter-rater agreement among the three graders was evaluated using Fleiss' kappa (k).

In the internal test set, PSNR and SSIM improved significantly from 23.2 and 0.823 (undilated) to 28.7 and 0.923 (model-enhanced) when compared to ground truth dilated images (both p < 0.001). The gradability of the images improved from 41% (49 of 120 images) to 76% after enhancement, although the inter-rater agreement decreased (overall Fleiss' κ from 0.397 to 0.234). Similar trends were observed in the external tests.

The enhanced images demonstrated improved quality and greater structural similarity to dilated images, suggesting the potential of the CofeNet model as an alternative approach to enhance image quality. Further validation is necessary to determine its clinical utility.