Transfer Learning using Macular OCTA to Differentiate LHON Vasculature from Normal Vasculature
First Prize, COS Awards of Excellence in Ophthalmic Research
Authors: Henry Liu, William Sultan, Kashif Iqbal, Valerio Carelli, Chiara LaMorgia, Patrick Yu-Wai Man, Piero Barboni, Alfredo Sadun, Rustum Karanjia.
Author Disclosure Block: H. Liu: None. W. Sultan: None. K. Iqbal: None. V. Carelli: None. C. LaMorgia: None. P. Man: None. P. Barboni: None. A. Sadun: None. R. Karanjia: None.
Purpose: Leber's Hereditary Optic Neuropathy (LHON) is the most common inherited mitochondrial disorder affecting complex I of the electron transport chain and leading to profound acute/subacute loss of central vision. In the acute phase of the disease, misdiagnosis is common. Recent advances in OCT angiography (OCTA) provides a multilayered image visualization of the choriocapillaris and choroid which can be segmented and analysed as three-dimensional voxels. While subtle changes of the vasculature around the optic nerve have been described in LHON, the retinal vasculature around the macula region is typically considered to be normal by clinical evaluation. The purpose of this study is to utilize a transfer learning neural network to differentiate between LHON and off-pedigree controls based on static macular OCTA images.
Study Design: Prospective case-control study Methods: 304 eyes (198 LHON and 106 Controls) underwent imaging with ZEISS PLEX Elite 9000 with AngioPlex Swept-Source OCT Angiography (ZEISS, Dublin, CA). Multiple Angioplex® 6mm x 6mm maps, centered on the macula, were acquired. Using the superficial layer, the center of the foveal avascular zone (FAZ) was chosen as the center of each image and pre-processing and standardization done using ImageJ. A deep-learning convolutional neural network (CNN) architecture, ResNet-152, was employed for this study. Transfer learning was implemented to retrain the CNN for optimized OCTA classification. Dataset was divided into 80% training set and 20% validation set. Five-fold cross validation technique was used to establish final accuracy for model performance.
Results: The pre-trained transfer learning algorithm was able to identify a subject as having a LHON gene mutation, when compared to controls with a five-fold cross-validation accuracy 93.3% with 100% sensitivity and 90.7% specificity, giving a positive predictive value of 81.0% and a negative predictive value of 100% with a Matthews correlation coefficient (MCC) of 0.857. The model was further validated using a separate data set of 48 eyes with normal tension glaucoma (NTG), which generate a similar optic neuropathy without clinically apparent change in the macula. In this preliminary validation model the algorithm failed to differentiate controls from NTG patients at a rate better than by chance using macular OCTA images. Yet, it was able to identify LHON asymptomatic carriers from NTG with a high degree of accuracy (100%).
Conclusions: Rare conditions, such as LHON, pose a diagnostic challenge for machine learning as large data sets are not readily available. By using a transfer learning model, we were able to use a smaller number of images than would classically be required for traditional deep learning. Our pre-trained deep learning algorithm was capable of correctly identifying patients who carry a LHON mutation from normal off-pedigree healthy controls based on static macular OCTA images alone. These findings suggest that the retinal vasculature of the macular region has peculiarities which are specific for LHON and is not a result of optic atrophy and can aid in the effective diagnosis of this debilitating inherited disease.