Abstract
This document serves as a formal correction to the article "From 2 dimensions to 3rd dimension: Quantitative prediction of anterior chamber depth from anterior segment photographs via deep-learning," originally published in PLOS Digital Health. The correction specifically addresses inaccuracies in the reporting of the testing dataset composition, ensuring the integrity of the demographic and ocular data used to validate the deep-learning model.
Introduction
A critical risk factor for primary angle-closure glaucoma is a shallow Anterior Chamber Depth (ACD)—the physical distance between the cornea and the iris. While measuring ACD typically requires expensive, specialized equipment like Optical Coherence Tomography (OCT), the original research proposed a deep-learning model capable of estimating this 3D metric from standard 2D anterior segment photographs. This correction notice provides the accurate distribution of participants within the validation cohorts used to test this model.
Datasets and Cohort Characteristics
The study utilized several population-based eye studies from Singapore to train and validate the regression model. The correction clarifies the specific participant counts for the testing phase, specifically within the Singapore Chinese Eye Study (SCES) and the Singapore Indian Eye Study (SINDI).
As shown in the corrected data, the testing dataset comprises:
- SCES (Singapore Chinese Eye Study): 229 participants (corrected from 257).
- SINDI (Singapore Indian Eye Study): 151 participants (corrected from 163).
- SiMES (Singapore Malay Eye Study): 943 participants.
- ISF (Iris Surface Features study): 343 participants.
While the total number of eyes in the testing set remains 380, the ethnic and cohort-specific distribution has been updated to reflect the true data structure.
Key Results and Insights
The corrected table provides essential "ground truth" measurements for the cohorts. For instance, the mean ACD for the SCES group was 2.56 mm (SD 0.37), while the SINDI group showed a mean of 2.61 mm (SD 0.34).
The inclusion of diverse cohorts (Malay, Chinese, and Indian) is vital for the development of generalizable AI tools in ophthalmology. Precise reporting of these demographics ensures that the model's performance is understood in the context of different ocular morphologies and ethnic backgrounds, which is a cornerstone of responsible medical AI research.
Conclusion
This correction reinforces the importance of data integrity and transparency in deep-learning research. By providing the accurate composition of the testing datasets, the authors ensure that the validation of their 2D-to-3D depth prediction model is reproducible and scientifically sound. Future work will continue to explore the utility of this low-cost screening tool in broader, more diverse global populations.
Summary updated