A deep learning model is able to develop objective and quantitative estimates of retinal nerve fiber layer (RNFL) thickness that correlate well with spectral-domain optical coherence tomography (SD-OCT) measurements, according to Duke University researchers.
Researchers conducted a retrospective cohort study to investigate whether RNFL thickness predictions attained using a deep learning model applied to fundus photographs can detect progressive glaucomatous changes over time. The research team collected 86,123 pairs of color fundus photographs and SD-OCT images taken during 21,232 patient visits. In total, images from 8831 eyes of 5529 patients with glaucoma or, from glaucoma suspects, were included. The team trained a deep learning convolutional neural network to assess fundus photographs and predict SD-OCT global RNFL thickness measurements. Then, the model was tested on an independent sample of eyes that had a longitudinal follow-up with both fundus photography and SD-OCT. Receiver operating characteristic (ROC) curves were used to assess the ability to detect eyes that had statistically significant slopes of SD-OCT change. The repeatability of RNFL thickness predictions was investigated by measurements obtained from multiple photographs acquired during the same day.
The test sample consisted of 33,466 pairs of fundus photographs and SD-OCT images collected from the 1147 eyes of 717 patients. Eyes in the test sample were followed up for an average of 5.3 years ± 3.3 years, with an average of 6.2 visits ± 3.8 visits. A significant correlation was found between change over time in predicted and observed RNFL thickness (r=0.76; 95% CI, 0.70-0.80; P <.001). RNFL showed a ROC curve area of 0.86 (95% CI, 0.83-0.88) to discriminate progressors from nonprogressors. For detecting fast progressors (slope faster than 2 mm/year), the ROC curve area was 0.96 (95% CI, 0.94-0.98), with a sensitivity of 97% for 80% specificity and 85% for 90% specificity. For photographs obtained at the same visit, the intraclass correlation coefficient was 0.946 (95% CI, 0.940-0.952), with a coefficient of variation of 3.2% (95% CI, 3.1%-3.3%).
Monitoring the progression of glaucoma helps identify patients who may require more aggressive treatment. While progression has traditionally been measured by assessing changes in visual field sensitivity, many patients with glaucoma show optic disc or RNFL changes without detectable deterioration on perimetric tests. Detecting structural changes provides an opportunity to start or increase treatment and help diagnose glaucoma in those only suspected of having the disease.
Because of its ability to quantify neural loss objectively, SD-OCT is widely used to assess longitudinal structural changes in patients with glaucoma. However, the technology is cost-prohibitive and requires well-trained technicians. As a result, it is not common in many developing countries. Recent advances in artificial intelligence algorithms — and deep learning neural networks, in particular — may provide a method of automating the assessment of structural glaucomatous damage using fundus photography, the researchers explain.
The study’s authors noted several limitations, including that the assessment of the study’s repeatability wasn’t planned prospectively, so researchers used an opportunistic sample of photographs that happened to have been obtained on the same visit day. Second, the team didn’t evaluate the test-retest reproducibility of photographs acquired on different days and by different photographers. Third, the machine-to-machine algorithm was trained to estimate only the global average RNFL thickness measurements. Lastly, the study would benefit from external validation in populations from different geographic areas, particularly areas with a high prevalence of myopia.
Disclosure: Several study authors declared affiliations with the biotech or pharmaceutical industries. Please see the original reference for a full list of authors’ disclosures.
Medeiros FA, Jammal AA, Mariottoni EB. Detection of progressive glaucomatous optic nerve damage on fundus photographs with deep learning. Ophthalmol. 2021;128(3):383-392. doi:10.1016/j.ophtha.2020.07.045