Accurately Identify nAMD Patients with Low Anti-VEGF Treatment Need by Deep Learning

Abstract

Purpose : We sought to identify nAMD patients with low anti-VEGF (aVEGF) treatment (Tx) need in the HARBOR study (NCT00891735) PRN arms utilizing Deep Learning (DL) algorithms.

Methods : We analyzed study eye OCTs captured during the 3 consecutive monthly loading ranibizumab (RBZ) injections (injs) from PRN arms of HARBOR, and correlated them with patient inj burden during the follow-up phase (3- to 23-month visit). Low aVEGF Tx need was defined as requiring ≤5 injs over 21 visits in the follow-up phase (Bogunovic H et al. Invest Ophthalmol Vis Sci. 2017;58(7):3240-3248). A study eye was eligible for analysis if Tx records were available for all visits in the follow-up phase.

OCT images (1024×512×128 voxels) acquired with Zeiss Cirrus machines were flattened towards the retinal pigment epithelium (RPE) layer and cropped with 384 pixels above (0.77mm), 128 pixels (0.46mm) below RPE. The central 15 B-scans were selected for model training and validation.

Our end-to-end DL model consisted of 10 convolution blocks to exponentially increase expressiveness and contained 5599 weight parameters to train fast. We used F-CNN architecture to enable patches of arbitrary size as input including the whole slice. Stochastic cropping was applied to sample 2D patches of random or prespecified size from B-scans. A committee machine was used for model ensemble. Predictions were aggregated from patch level to the patient level to magnify weak signal to strong signal. Stratified 5-folds were created at the patient level and nested cross-validation was used for performance evaluation.

Results : Among 547 PRN study eyes, 352 were eligible for analysis. Among these, 79 (22.4%) eyes were classified as low aVEGF Tx need. Our model achieved area under the receiver-operating curve (AUROC) 78.6% with 95% confidence interval [72.7, 84.4%] on the validation set.

Conclusions : Our study showed that our prototype DL model had high performance predicting nAMD patients with low aVEGF Tx need after the RBZ loading phase. Our findings are consistent with other studies and results are better than those previously reported. While this algorithm needs to be validated on other databases and in the real-world setting, it could help physicians/patients understand future Tx burden. It could also stratify patients in future clinical trials assessing aVEGF inj durability.

This is a 2020 ARVO Annual Meeting abstract.

 

Fig 1. DL pipeline for training/validation

 

Fig 2. ROC for model performance evaluation