Abstract
Purpose :
Grading the aperture of the iridocorneal angle (ICA) is fundamental for the categorization of glaucoma and for treating related conditions. As this task requires specific skills and may be time-consuming, we devised a new automatic system as a supporting tool for ICA aperture estimation in digital gonio-photographs.
Methods :
609 ICA sector images acquired by a NIDEK GS-1 device (NIDEK CO., LTD. Japan) were each annotated by grading the ICA aperture in 60 points sampled along the interface. 3 aperture grades based on a simplified Spaeth’s scale were considered: Open, i.e., Spaeth’s grades D and E, Occludable, i.e., grade C, and Closed, i.e., grades A and B. The overall number of per-class samples was balanced.
A custom deep-learning (DL) model was designed and trained on this dataset to interpret anatomical features of the ICA and map input shots into densely sampled vectors of local aperture estimates. Our approach differs from previous research since we consider ICA aperture estimation as a regression rather than a classification task, allowing to generate detailed aperture “profiles” instead of just labels, inadequate to properly describe the variability of the morphology in large ICA regions.
By randomly changing the model’s structure (i.e., Monte Carlo dropout) we generate multiple ICA profile candidates and use their sample-wise variability as a proxy for results uncertainty to identify less robust estimates.
Results :
The algorithm’s average performance was evaluated through a 5-fold cross-validation experiment returning a per-sample accuracy of about 70%. The model’s sensitivity for all three classes was comprised between 65% and 68%. Precision and specificity equaled 73% and 86% for the Open class, and 79% and 92% for the Closed class. The Occludable class returned lower values (54% and 71%) as choosing the middle grade is less costly in a regression problem when the input is not well characterized, thus generating more false positives.
Conclusions :
ICA aperture profiles estimated by our model, despite being trained on a limited pilot dataset, show good agreement with ground truth. Uncertainty estimation improves results’ interpretability. The integration of this algorithm in an advanced tool for exam visualization and annotation may ease and speed-up the analysis of digital gonio-photographs making the thorough assessment of the ICA available also to non-specialized professionals.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.