Self-supervised machine learning for individual prediction of conversion to neovascular AMD in PINNACLE study

Abstract

Purpose : The lack of well-established biomarkers and a wide variability in the progression speed makes predicting the conversion from intermediate to neovascular age-related macular degeneration (iAMD, nAMD) a challenging task. We developed an artificial intelligence (AI) method to predict the risk of conversion of an eye from iAMD to nAMD within the next 6,12 and 18 months from OCT scans.

Methods : We propose a two-stage Deep Learning method. First, a fully convolutional Encoder is learned for feature extraction through Self-Supervised learning (SSL) to overcome the paucity of annotated data. Next, a 3 layer Classifier is trained to predict the probability of the time to conversion P(T*≤t) from the learned features (Fig. a).
The Encoder is trained with a novel SSL task using pairs of unlabelled OCT scans of the same eye from two different visits. The change between the features extracted from two visits should reflect the structural deformation and intensity changes between them. To achieve this, we jointly train the Encoder with a Decoder Network (Fig. b) to predict the transformation that morphs the OCT scan of the first visit to the second using the difference between their features as input.
We model P(T*≤t) as a sigmoidal distribution (Fig. c) whose parameters are predicted by a Classifier trained in a supervised manner.

Results : We analyzed a subset of the Fellow Eyes from the Retrospective cohort of the PINNACLE trial. SSL was trained on Topcon Scans of 399 eyes (3570 visits) without manual annotations. Stage 2 employed a five-fold cross-validation for evaluation on 2418 OCT volumes from 343 eyes with manual labels of the conversion date. Our SSL pre-trained method achieved an average Area under the ROC curve of 0.766 in predicting the conversion of eyes within the next 6 months from the input Scan, while a fully supervised trained model achieved 0.71 (Table).

Conclusions : We proposed an AI system with a novel SSL method for capturing the temporal changes in OCTs. Our automated method to predict the future risk of the onset of nAMD can play a critical role in enabling patient-specific disease management and enhancing clinical trial populations with patients at risk.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Our 2-stage method in (a) comprises an Encoder trained via SSL in (b), and a classifier to predict P(T*≤t) in (c)

 

Our performance compared to other 3D CNN (I3D, X3D) and SSL methods (Model Genesis, time interval prediction)