AI algorithm assisted anti-VEGF therapy in neovascular AMD

Author: Daniel Pauleikhoff (Germany)

Co-authors: Matthias Gutfleisch, Peter Mussinghoff, Kai Rothaus, Martin Quassowski, Oliver Ester

Purpose

Anti-VEGF treatment in neovascular AMD (nAMD) has been established as “gold standard” reducing visual loss significantly. Therefore this therapy is widely introduced in the care of these patients. Especially in Germany cooperative therapeutical system including general ophthalmologists and retina specialists have been develop. To support general ophthalmologists to refer and monitor AMD patients during this ongoing cooperative treatment, AI algorithm and assistance may of important. Also AI may support counseling of patients in respect to the prediction of clinical relevant parameters like length and intensity of treatment.

Setting/Venue

SD-OCT and clinical data of an unselected consecutive real-life cohort of 1347 AMD patients, treated between 2014-20 (IVAN therapy strategy) and with longterm follow-up were analysed.

Methods

To support referral for treatment and monitoring retreatment indications on SD-OCTs a novel deep learning network (DLN) based on SD-OCT volume scans (49 scans) was developed differentiating between early/intermediate vs nAMD (referral) and between stabilized vs retreatment nAMD (retreatment). In addition improved image processing and mixed data analysis were used to predict the individual treatment intensity (number of treatment IVAN cycles) in the first 12 mo after the loading phase. In a step towards visualisation of the AI results, saliency maps of the SD-OCT volume scans were generated.

Results

The DLN algorithm to differentiate between early/intermediate vs nAMD SD-OCT achieved an AUC of 0.932 (10-fold CV-AUC) and between stabilized vs retreatment nAMD SD-OCT an AUC of 0,865 (10-fold CV-AUC). The best DLN algorithm to predict the number of future injection cycle (1 vs 3 IVAN-cycle) in the next 365 days after upload included the initial SD-OCT volume scans and the time until the first visit with renewed lesion activity. With this experimental structure an AUC of 0.71 (10-fold CV-AUC) was achieved with an AUC of 0.84 only including eyes with a trust-filter >0.6.

Conlusions

Assistance by the new developed AI algorithms can support ophthalmologists at first in their initial decision to refer AMD patients to retina specialists for anti-VEGF therapy, but secondly to monitor patients during the course of the therapy. For both the developed AI algorithm demonstrated a very high AUC and accurancy. In addition further AI algorithm may help to predict the future intensity of anti-VEGF-therapy. Saliency maps of the individual SD-OCT volume scans may help ophthalmologists and patients to understand the individual decisions of the AI algorithm.

Financial Disclosure

The study was supported by a grant of Novartis