Deep learning to infer visual acuity from optical coherence tomography in diabetic macular edema

Diabetic macular edema (DME) is one of the leading causes of visual impairment in diabetic retinopathy (DR). Physicians rely on optical coherence tomography (OCT) and baseline visual acuity (VA) to tailor therapeutic regimen. However, best-corrected visual acuity (BCVA) from chart-based examinations may not wholly reflect DME status. Chart-based examinations are subjected findings dependent on the patient's recognition functions and are often confounded by concurrent corneal, lens, retinal, optic nerve, or extraocular disorders. The ability to infer VA from objective optical coherence tomography (OCT) images provides the predicted VA from objective macular structures directly and a better understanding of diabetic macular health. Deviations from chart-based and artificial intelligence (AI) image-based VA will prompt physicians to assess other ocular abnormalities affecting the patients VA and whether pursuing anti-VEGF treatment will likely yield increment in VA.

We enrolled a retrospective cohort of 251 DME patients from Big Data Center (BDC) of Taipei Veteran General Hospital (TVGH) from February 2011 and August 2019. A total of 3,920 OCT images, labeled as "visually impaired" or "adequate" according to baseline VA, were grouped into training (2,826), validation (779), and testing cohort (315). We applied confusion matrix and receiver operating characteristic (ROC) curve to evaluate the performance.

We developed an OCT-based convolutional neuronal network (CNN) model that could classify two VA classes by the threshold of 0.50 (decimal notation) with an accuracy of 75.9%, a sensitivity of 78.9%, and an area under the ROC curve of 80.1% on the testing cohort.

This study demonstrated the feasibility of inferring VA from routine objective retinal images.

Serves as a pilot study to encourage further use of deep learning in deriving functional outcomes and secondary surrogate endpoints for retinal diseases.