Researchers developed a deep learning model that outperforms conventional approaches for predicting visual function (VF) from optical coherence tomography images, according to a study published in the American Journal of Ophthalmology.
The researchers aimed to develop new diagnostic technology using spectral domain optical coherence tomography (SD-OCT) images and retinal nerve fiber layer thickness (RNFLT) measurements. The main outcome measures were pointwise prediction mean error, mean absolute error, and correlation of predictions with the best available estimate (BAE) VF sensitivity.
They developed 2 standard reference models and 2 deep learning models that predict pointwise VF sensitivity. Among the deep learning models, model 1 exclusively used the RNFLT profile, while model 2 used the RNFLT profile and SD-OCT images. They used an independent test-retest dataset consisting of 2181 SD-OCT/visual function pairs from 954 eyes of participants, with and without glaucoma (n = 332 and n = 220, respectively), to test all models. Median values of approximately 10 VFs per eye were used to determine the BAE of the true VF. The team also evaluated the performance of single VFs in predicting the BAE VF and evaluated the models on an independent test dataset derived from eyes of participants with glaucoma (n = 72).
The median mean deviation of the VF from standard automated perimetry was -4.17 decibels (dB) among glaucomatous eyes (5th-95th percentile, -14.22-0.88dB). According to the investigators, model 2 had excellent accuracy, with a mean error of 0.5 dB (standard deviation (SD): 0.8). For overall performance, it had a BAE pointwise prediction mean absolute error of 2.3 dB (SD: 3.1), significantly outperforming all other models.
The pointwise mean absolute error was 1.5 dB (SD: 0.7) when model 2 was used with single VFs to predict BAE VF. The researchers also documented favorable associations between SD-OCT and single VF predictions of the BAE pointwise VF sensitivities (R2 = 0.78 and R2 = 0.88, respectively).
“We have introduced a methodology for translating functional and structural measurements used in the clinical evaluation of glaucoma into the same domain – predicting the VF from OCT images,” according to the investigators. “Estimates of functional deficits from structural measures yielded from this method are better than those derived from previous approaches and approach the accuracy of single VF tests.”
Study limitations include the possibility of a floor effect, with respect to glaucoma severity, leading to overestimation at the lower end of VF sensitivity.
Disclosure: This research was supported by Santen Pharmaceutical Co., Ltd. Some study authors declared affiliations with the biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of disclosures.
Lazaridis G, Montesano G, Afgeh SS, et al. Predicting visual fields from optical coherence tomography via an ensemble of deep representation learners. Am J Ophthalmol. Published online January 4, 2022. doi:10.1016/j.ajo.2021.12.020