Abstract
Purpose :
Deflection of blood vessels over the neuroretinal rim serve as a biomarker for neuroretinal rim loss in glaucomatous optic neuropathy. Retinal vasculature patterns have been proposed as alternative or adjunctive biomakers to optic disc evaluation to establish glaucoma diagnosis in fundus images. The purpose of the present work is to compare human expertise and deep learning algorithms in the ability to detect glaucoma through assessment of retinal vasculature alone.
Methods :
The deep learning algorithm was trained using 437 retinal vasculature maps (265 non-glaucomatous, 172 glaucomatous) from 4 publicly available labeled datasets. 4 different architectures of a convolutional neural network model were examined with convolutional layers varying from 4 to 1 reducing parameter size respectively. 10 images from each dataset (40 images total) were randomly selected for evaluation and validation by two expert clinical graders. All available metadata was removed and graders were masked to the labeled reference standard. Graders independently evaluated each image to determine the presence or absence of glaucoma. Agreement between graders and agreement of graders with labeled ground truth for each image was determined by percent agreement.
Results :
Agreement of graders with ground truth labels was 60% for grader one and 67.5% for grader two, which improved to 77.5% following discussion. Agreement with labels was poor in instances where the level of image detail was low. The deep learning model was not able to determine the presence of glaucoma based on retinal vasculature maps alone and the architectures performed 50%±1%, which is not greater than chance for the binary classification despite data augmentation, model size reduction, and convolution depth reduction.
Conclusions :
Agreement of presence or absence of glaucoma solely based on retinal vasculature patterns between expert graders was moderate, while a deep learning model was not able to determine the presence or absence of disease. Future work will increase sample size by orders of magnitude and explore transfer learning from foundation models in order to improve model performance.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.