Deep learning method for semi-automated segmentation of optic nerve head tissues in optical coherence tomography images

Optical coherence tomography (OCT) enables acquisition of image volumes of the optic nerve head (ONH) in vivo. These image volumes provide structural and biomechanical information about the ONH. The individual tissues in these images must be segmented to obtain tissue-specific information, but doing so manually is time consuming. We present a method which uses a deep learning model to semi-automatically segment the following tissue boundaries of the ONH: the anterior lamina cribrosa surface (ALCS), Bruch's membrane (BM), and choroid-scleral interface (CS). We trained a convolutional neural network (CNN) to predict tissue boundary locations using 46 pre-processed ONH image volumes consisting of 24 radial scans each and their corresponding manual segmentations. When the trained model is presented with a new image, a user is able to accept the model's predictions or drag any misplaced markings to their correct locations. These corrected or accepted markings are used to update the model's predictions for subsequent images. The root mean squared error (RMSE) was used to evaluate the difference between the updated model's predictions and the manually marked segmentations. The RMSE was found to be 8.98 ± 14.74 μm for the BM, 24.26 ± 17.96 μm for the CS, and 49.69 ± 32.08 μm for the ALCS. Updating the model improved predictions for the BM and CS for our test set consisting of image volumes of 6 ONHs, with 24 scans per ONH. A more complete evaluation of the updated model's effect on prediction accuracy could be obtained using a larger test dataset.