Detection of Progressive Glaucomatous Optic Nerve Damage on Fundus Photographs with Deep Learning

Medeiros, Felipe A.; Jammal, Alessandro A.; Mariottoni, Eduardo B.

PURPOSE: To investigate whether predictions of retinal nerve fiber layer (RNFL) thickness obtained from a deep learning model applied to fundus photographs can detect progressive glaucomatous changes over time. DESIGN: Retrospective cohort study PARTICIPANTS: 86,123 pairs of color fundus photos and spectral-domain optical coherence tomography (SD OCT) scans collected over 21,232 visits from 8,831 eyes of 5,529 patients with glaucoma or suspects. METHODS: A deep learning convolutional neural network was trained to assess fundus photographs and predict SD OCT global RNFL thickness measurements. The model was then tested on an independent sample of eyes that had longitudinal follow-up with both fundus photos and SD OCT scans. The ability to detect eyes that had statistically significant slopes of SD OCT change was assessed by receiver operating characteristic (ROC) curves. The repeatability of RNFL thickness predictions was investigated by measurements obtained from multiple photos that had been acquired during the same day. MAIN OUTCOME MEASURES: The relationship between change in predicted RNFL thickness from photos and change in SD OCT RNFL thickness over time. RESULTS: The test sample consisted of 33,466 pairs of fundus photos and SD OCTs collected over 7,125 visits from 1,147 eyes of 717 subjects. Eyes in the test sample were followed for an average of 5.3 ± 3.3 years, with an average of 6.2 ± 3.8 visits. There was a significant correlation between change over time in predicted and observed RNFL thickness (r = 0.76; 95% CI: 0.70 – 0.80; P < 0.001). RNFL predictions had ROC curve area of 0.86 (95% CI: 0.83 – 0.88) to discriminate progressors from non-progressors. For detecting fast progressors (slope faster than 2μm/year), the ROC curve area was 0.96 (95% CI: 0.94 – 0.98) with sensitivity of 97% for 80% specificity, and 85% for 90% specificity. For photos obtained at the same visit, the intraclass correlation coefficient was 0.946 (95% CI: 0.940 – 0.952), with coefficient of variation of 3.2% (95% CI: 3.1% - 3.3%). CONCLUSION: A deep learning model was able to obtain objective and quantitative estimates of RNFL thickness that correlated well with SD OCT measurements and could potentially be used to monitor for glaucomatous changes over time.