Multi-Comparison of Different Ocular Imaging Modality-based Deep Learning Models for Visually Significant Cataract Detection.

PURPOSE: Age-related cataract is the leading cause of vision impairment. Researchers have utilized various imaging modalities, including slit beam, diffuse anterior segment, and retinal imaging, to develop deep learning (DL) algorithms for automated cataract analysis. However, the comparative performance of these algorithms across different ocular imaging modalities remains unevaluated, mainly due to the absence of standardized test sets across studies.

DESIGN

Retrospective study.

PARTICIPANTS

Across all the models, the Singapore Malay Eye Study data set was used for training (N = 7093 eyes) and internal testing (N = 1649 eyes). The Singapore Indian Eye Study (SINDI; N = 5579 eyes) and the Singapore Chinese Eye Study (SCES; N = 5658 eyes) were used for external testing. A community study data set of nonmydriatic retinal photos (N = 310 eyes) was used for external testing of the retinal model.

METHODS

We developed 3 single-modality DL models (retinal, slit beam, and diffuse anterior segment photos) and 4 ensemble models (4 different combinations of the 3 single-modality models) to detect visually significant cataract (VSC). We defined eyes with VSC as having significant cataract (based on the modified Wisconsin cataract grading system) with a best-corrected visual acuity of <20/60.

MAIN OUTCOME MEASURES: Area under receiver operating characteristic curve (AUC).

RESULTS

In the internal test, the retinal model had the highest AUC value (97.0%; 95% confidence interval [CI], 95.9-98.2), compared with the slit beam model (AUC, 93.4%; 95% CI, 90.1-96.7; P diff = .029) and diffuse anterior segment model (AUC, 94.4; 95% CI, 92.3-96.4; P diff = .002). There was no significant difference in AUC when comparing the retinal model with the ensemble models (all P diff ≥ .07). These trends were consistently observed in the external test sets. In nonmydriatic eyes, the retinal model showed reasonable performance (AUC, 89.8%; 95% CI, 89.6-89.9).

CONCLUSIONS

Our findings highlight the retinal model as a promising tool for detecting VSC, outperforming slit beam and diffuse anterior segment models. Because retinal photography is routine in diabetic retinopathy screening, this approach could enable opportunistic cataract screening with minimal add-on cost.

FINANCIAL DISCLOSURE: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.