Personalizing Circumpapillary Retinal Nerve Fiber Layer Thickness Norms for Glaucoma

点击翻译按钮获取中文摘要

Project Summary Motivation and Hypotheses: The circumpapillary RNFL thickness (cpRNFLT) measured by circle scan is routinely used for glaucoma diagnosis. Precise cpRNFLT norms are important for assessing cpRNFLT abnormalities, while current optical coherence tomography (OCT) devices used in glaucoma care only adjust the cpRNFLT norms for age. Prior studies attempted to adjust cpRNFLT norms for retinal anatomy either by manually delineated features such as blood vessel location and disc-fovea angle or standard clinical metrics such as scan diameter and axial length, while manual feature extraction is laborious and standard clinical metrics are insufficient to represent the complex retinal anatomical variation. We hypothesize that we can leverage artificial intelligence (AI) modeling to (1) improve cpRNFLT norms by automatically adjusting for retinal anatomy encoded by retinal imaging data, which can be then used to (2) improve glaucoma diagnosis. Aim 1: Developing AI-based models to personalize cpRNFLT norms with individual retinal anatomy. Healthy subject data from the Leipzig population-based study will be used to develop Lasso linear regression and deep learning models to adjust pointwise cpRNFLT norms for retinal anatomy represented by inner limiting membrane (ILM) maps and scanning laser ophthalmoscopy (SLO) fundus images. 60%, 20% and 20% of the entire dataset will be used for training, validation and testing, respectively. The cpRNFLT norm accuracy will be measured by mean absolute error and R2. For the Lasso model, we will apply principal component analysis followed by uniform manifold approximation and projection to extract retinal anatomical features from the ILM map and SLO fundus image. For the deep learning model, we will use both the pre-trained deep learning model ResNet-50 and a custom designed convolutional neural network ignoring missing imaging values. Aim 2: Clinical relevance validation for the personalized cpRNFLT norms based on individual retinal anatomy. Glaucoma patient data from Massachusetts Eye and Ear will be used to demonstrate the clinical relevance of our personalized cpRNFLT norms with Lasso linear regression and deep learning models. The pointwise cpRNFLT deviation percentiles will be used to predict accompanying VFs. Mean absolute error and R2 on the testing subset will be used to evaluate model performance. Paired t-test will be performed to compare if using cpRNFLT deviation percentiles normalized by our personalized cpRNFLT norms can better predict VFs compared with by the standard cpRNFLT norms only adjusting for age, gender and scan diameter. For the deep learning model, a 1D convolutional neural network enhanced by attention units will be developed. Main Deliverables and Public Health Impacts: This project will construct personalized cpRNFLT norms by automatically adjusting for individual retinal anatomy using retinal imaging data with cutting edge AI technology. The success of this project may have a great impact to improve clinical care for glaucoma patients.