Application of a Deep Machine Learning Model for Automatic Measurement of EZ Width in SD-OCT Images of RP

PURPOSE: We applied a deep convolutional neural network model for automatic identification of ellipsoid zone (EZ) in spectral domain optical coherence tomography B-scans of retinitis pigmentosa (RP). METHODS: Midline B-scans having visible EZ from 220 patients with RP and 20 normal subjects were manually segmented for inner limiting membrane, inner nuclear layer, EZ, retinal pigment epithelium, and Bruch's membrane. A total of 2.87 million labeled image patches (33 × 33 pixels) extracted from 480 B-scans were used for training a convolutional neural network model implemented in MATLAB. B-scans from a separate group of 80 patients with RP were used for testing the model. A local connected area searching algorithm was developed to process the model output for reconstructing layer boundaries. Correlation and Bland-Altman analyses were conducted to compare EZ width measured by the model to those by manual segmentation. RESULTS: The accuracy of the trained model to identify inner limiting membrane, inner nuclear layer, EZ, retinal pigment epithelium, and Bruch's membrane patches in the test dataset was 98%, 89%, 91%, 94%, and 96%, respectively. The EZ width measured by the model was highly correlated with that by two graders (r = 0.97; P < 0.0001). Bland-Altman analysis revealed a mean EZ width difference of 0.30 mm (coefficient of repeatability = 0.9 mm) between the model and the graders, comparable to the mean difference of 0.34mm (coefficient of repeatability = 0.8 mm) between two graders. CONCLUSIONS: The results demonstrated the capability of a deep machine learning-based method for automatic identification of EZ in RP, suggesting that the method can be used to quantify structural deficits in RP for detecting disease progression and for evaluating treatment effect. TRANSLATIONAL RELEVANCE: A deep machine learning model has the potential to replace humans for grading spectral domain optical coherence tomography images in RP.