Purchase this article with an account.
Atalie C. Thompson, Alessandro A Jammal, Eduardo Bicalho Mariottoni, Leonardo Shigueoka, Tais Estrela, Felipe A Medeiros; Predicting Future Rates of Retinal Nerve Fiber Layer Loss from Deep Learning Assessment of Baseline Optic Disc Photographs. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4533.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To investigate the ability of a deep learning (DL) model trained on baseline optic disc (OD) photographs to predict subsequent rates of loss in retinal nerve fiber layer (RNFL) thickness in glaucoma and glaucoma suspects (GS), as measured by spectral-domain optical coherence tomography (SDOCT).
The study included 2,253 eyes of 1,294 patients with glaucoma or suspected of disease followed for an average of 3.5 ±1.9 years (range: 1.0 to 8.8 years). Data was extracted from the Duke Glaucoma Registry (DGR). All eyes had baseline OD photographs and were followed over time with SDOCT RNFL thickness measurements. Linear mixed models were used to estimate the rates of change in global RNFL thickness over time. A DL convolutional neural network (ResNet) was trained to predict the rates of RNFL change from the baseline image. Only a single baseline image was used per eye. The sample was divided into 80% for training/validation and 20% for final testing, with randomization at the patient level.
The test sample included 457 eyes with an actual mean rate of RNFL change on SDOCT of -0.75 ±1.6 mm/y, ranging from -9.57 to 1.84mm/y. The mean DL-predicted rate of RNFL change was -1.02 ±1.63 mm/y, ranging from 1.35 to -9.52mm/y. There was a significant association between the DL predictions of the rate of RNFL change from the baseline image and the actual rates of RNFL change on SDOCT (r = 0.43, R2= 19; P<0.001) (Figure). For comparison, baseline global RNFL thickness from SDOCT had a much weaker association with the actual rates of RNFL change over time (r=-0.09, R2=0.01; P=0.048), and there was a statistically significant difference between the R2 values (P<0.001). The areas under the receiver operating characteristic curves to discriminate fast (faster than -2mm/y) versus slow (slower than -0.5mm/y) progression were 0.78 for the baseline DL model’s assessment of OD photographs and 0.56 for baseline global RNFL thickness.
A DL model trained on baseline OD photographs was able to predict future rates of RNFL loss in glaucoma. The DL model’s predictions from these images performed better than baseline RNFL thickness, suggesting that there may be features of the optic disc indicative of risk for fast progression that are not fully conveyed by baseline SDOCT RNFL thickness measurements alone.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only