Enhanced RNFL Thickness Estimation with Cost Function Approach from Fundus Images via TSNIT Graph Mapping

This study endeavors to improve more efficient estimation of RNFL thickness by transforming fundus images into optical coherence tomography (OCT)-equivalent retinal nerve fiber layer (RNFL) maps. The ground truth is derived from TSNIT graphs, and the RNFL thickness maps are obtained from OCT reports, as there are limitations in accessing raw OCT data. We introduce a novel cost function within the autoencoder that considers both global and local thickness around the RNFL thickness map, integrated with TSNIT graphs for accurate estimation. Analyzing a dataset of 443 fundus images across three groups–103 glaucoma, 71 glaucoma suspected, and 269 no glaucoma–we observed the highest micro-average values for peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) at 24.92 dB and 91.14%, respectively. Examining individual groups, consistent PSNR and SSIM measurements were evident: glaucoma registered 26.26 dB and 92.30%, glaucoma suspected exhibited 25.07 dB and 91.91%, and the no glaucoma group demonstrated 24.41 dB and 90.73%. This demonstrated improved RNFL thickness estimates and created more accurate TSNIT graphs using an innovative method based solely on fundus images.