Developing Population-Generalizable Deep Learning Models for Automated Glaucoma Screening

Schepens Eye Research Institute, Boston MA

Investigators

Linked publications & trials

Abstract

Project Summary Motivation: Glaucoma is a leading cause of irreversible blindness affecting 3 million patients in the US. About 50% of people with glaucoma are unaware of their condition, particularly due to a lack of access to routine ophthalmic care. Glaucoma detection with deep learning using retinal imaging is promising to provide cost-effective mass glaucoma screening, which can be deployed in primary care and pharmacies without needing individuals to visit more expensive eye clinics for screening. It is unclear if deep learning models for glaucoma detection are population-generalizable, and few studies have been conducted to develop novel methods to improve model population generalizability. We hypothesize that performance consistency across populations in glaucoma detection can be improved by developing novel population-generalizable deep learning techniques. We propose new generalizability-scaled performance metrics of the area under the receiver operating characteristic curve, sensitivity, and specificity for population generalizability assessment, which penalize raw performance metrics with performance variance across populations. Overall Approach: Our models use optical coherence tomography scans and fundus photos to predict glaucoma diagnosis defined by clinical guidelines. Our population-generalizable models will be compared with standard models with and without oversampling and transfer learning measured by generalizability-scaled performance metrics. Aims: (1) Assessing performance generalizability of existing deep learning models and developing a population scaling approach. We will quantify the performance population generalizability for the three widely used deep learning models including ResNet, EfficientNet, and Vision Transformer. We will develop a novel population scaling approach to improve model generalizability by reweighting the training error function based on the cross-population variance of training errors. (2) Developing population-conditioned generative models for data augmentation to improve performance generalizability. We will develop population-conditioned generative models to synthesize training data. The generated data will be used to improve data availability across populations for training glaucoma detection models. (3) Developing population normalization models to improve glaucoma detection generalizability. We will develop a population normalization technique to balance feature importance across populations. We will integrate the three population-generalizable approaches to maximize performance generalizability. Main Deliverables and Impacts: Our population-generalizable deep learning models for automated glaucoma screening will greatly benefit patients by providing more effective glaucoma screening. We will release a large de-identified dataset of 10,000 patients to the public to study the population generalizability of medical artificial intelligence.