Abstract
Purpose :
A deep learning-based approach to monitoring pediatric retinal disease may be limited by the size of the available data set due to the challenges of image collection in a young and vulnerable patient population. This study investigated a potential data augmentation approach to address data scarcity by translating rodent infrared reflectance (IR) fundus images, focusing on vessel patterns, into human IR images using generative artificial intelligence (AI).
Methods :
Anesthetized Long-Evans rats underwent routine imaging sessions after pupil dilatation. IR fundus images (n = 36) were acquired using the Heidelberg Spectralis. Human IR fundus images (n = 36) were sourced from the publicly available Retinal Arteries and Veins in Infrared Reflectance imaging (RAVIR) dataset obtained with the Spectralis. Small patches (768 by 768 pixels) of rat retinal vessels (excluding major vessels) were used to 'translate' into the human fundus. A generative AI model using a cycle generative adversarial network was trained for image translation. Three graders with different levels of expertise evaluated the test set, which comprised the translated IR image patches (n = 26) randomly interspersed with human IR image patches (n = 26). Three graders were blinded to the ground truth and were asked to independently determine whether the images originated from human subjects or were translated.
Results :
The overall accuracy of the grader's assessment is 0.84 ± 0.11 in this pilot study. A high standard deviation (0.11) indicated that the translated images retained similar vascular patterns of human IR images to create uncertainties for the graders. It is further supported by an intraclass correlation coefficient of 0.59, a moderate positive correlation among the graders. The graders could distinguish translated images from human IR images most of the time instead of randomly guessing due to indistinguishable features (yielding a value of 0.5). It is worth noting that such accuracy may be influenced by graders gaining extra information using normal rat vessel patterns for translating, but a human dataset comprises seven different retinal dysfunctions.
Conclusions :
This study demonstrated the feasibility of translating rodent IR images into human IR images, focusing on vasculature patterns. This approach may be useful in data augmentation when data scarcity exists.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.