Abstract
Purpose :
A major limitation in managing age-related macular degeneration (AMD) is that we can only estimate the average risk of conversion from dry to neovascular AMD (NVMAD) over 5 years based on AREDS criteria. Our inability in identifying patients at high risk of imminent conversion limits our ability to initiate treatment at the earliest sign of exudation, and it is known that better presenting visual acuity (VA) predicts better final VA. The current study aims to analyze 3D optical coherence tomography (OCT) volumes using deep learning (DL) and to create a model that can provide more fine-grained prediction for imminent conversion (within 6 months) from dry to NVAMD.
Methods :
Inclusion criteria: patients with one or both eyes that have converted from dry to NVAMD. Exclusion criteria: concurrent significant retinal pathologies other than AMD and unavailability of OCT data. Transfer learning and data augmentation were used to train a 3D neural network (MedicalNet-50) for the binary classification task of distinguishing between 3D OCT volumes obtained more than 6 months prior to vs. within 6 months of conversion. The conversion date was defined as the 1st anti-VEGF injection date. The model’s performance was evaluated by 5-fold cross validation, and data partition was performed on a patient level.
Results :
5,440 OCT volumes from 1,842 patients were included. Using a single OCT volume as input, our model achieved the following average performance during 5-fold cross validation in predicting imminent conversion to NVAMD within 6 months: area under the receiver operating characteristic curve (AUC) 0.75±0.04, sensitivity 0.81±0.04, and specificity 0.63±0.05.
Conclusions :
Using only a single OCT volume as input, our DL-based model was able to predict with reasonable performance imminent conversion to NVAMD within 6 months. Human retinal specialists are typically incapable of providing such predictions in a consistent manner, and 6 months is a more clinically meaningful time frame as compared to the 5 years (the duration over which average risk can be estimated using AREDS criteria). Being able to identify patients at a high risk for imminent conversion will enable more personalized management (e.g. more frequent clinic visits and prescription of at-home OCT devices), and could lead to more timely initiation of treatment and better long-term VA outcome.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.