Deep Learning–based Estimation of Glomerular Filtration Rate from Macular Optical Coherence Tomography
10.21561/jor.2025.10.2.182
- Author:
Kun-Hoo NA
1
;
Yong Sung YOU
Author Information
1. Nune Eye Hospital, Seoul, Korea
- Publication Type:ORIGINAL ARTICLE
- From:
Journal of Retina
2025;10(2):182-189
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:This study aimed to develop a deep learning regression model using macular optical coherence tomography (OCT) images to predict estimated glomerular filtration rate (eGFR).
Methods:In this retrospective cross-sectional study, we analyzed data from patients who underwent preoperative evaluation for cataract surgery, including spectral-domain OCT (Spectralis; Heidelberg Engineering) images, blood test–derived eGFR, and ocular biometry.A dual-input regression model based on a modified ResNet-18 architecture was constructed to process simultaneously horizontal and vertical B-scan macular OCT images. The model’s predicted eGFR was further combined with patient age, sex, presence of diabetes and hypertension, and axial length using a Random Forest regression algorithm to create a hybrid model. Model performance was assessed using the root mean squared error (RMSE) and coefficient of determination (R2 ). Gradient-weighted regression activation mapping (GradRAM) was applied for visualization to assess the anatomical regions contributing to predictions.
Results:A total of 101 eyes from 101 patients were included. The ResNet-18 model achieved an eGFR prediction accuracy with an R2of 0.89 and an RMSE of 4.23. The Random Forest hybrid model further improved predictive performance, achieving an R2 of 0.94 and an RMSE of 3.64. The mean absolute error (MAE) across the full cohort was 3.69 ± 2.08 mL/min/1.73 m2 . Among the 10 most accurately predicted cases, the MAE was as low as 0.42 ± 0.20 mL/min/1.73 m2 , while the 10 least accurate cases showed an MAE of 7.44 ± 0.92 mL/ min/1.73 m2 . Grad-RAM visualizations revealed predominant activation in the choroidal region of vertical OCT images.
Conclusions:This is the first study to demonstrate accurate prediction of renal function from macular OCT images using deep learning.The dual-input architecture and Grad-RAM visualization enabled high prediction performance and partial anatomical interpretability.These findings support the potential of OCT-based oculomics for systemic disease assessment.
