Name

Deep Learning pour la prédiction automatique de l’activation précoce des membranes néovasculaires quiescentes naïves de traitement dans le contexte d’une DMLA

Around 30% of quiescent macular neovascularizations (qMNVs) exudate within 2 years of follow-up in patients with age-related macular degeneration (AMD). The aim of the study is to develop a deep learning classifier based on optical coherence tomography (OCT) and OCT angiography (OCTA) images to automatically identify qMNVs at risk for early exudation.

AMD patients showing OCTA-documented qMNV with a 2-years minimum imaging follow-up were retrospectively selected from two specialized retina centers. Patients showing OCT B-scan-documented MNV exudation within the first 2 years formed the EX GROUP while the others formed the QU GROUP. ResNet-101, Inception-ResNet-v2 and DenseNet-201 were independently trained first on OCTA images and secondarily on OCT B-scan images. The resulting 6 models were merged using major and soft voting techniques to increase classification performance.

Eighty-nine patients with a follow-up of 5.8 ± 1.7 years were recruited (35 in the EX GROUP and 54 in the QU GROUP). Inception-ResNet-v2 was the best performing among the 3 single convolutional neural networks (CNNs). The major voting model resulting from the association of the 3 different CNNs resulted in improvement of performance both for OCTA and OCT B-scan, with 89.3% and 90.0% testing accuracy respectively. Soft voting model resulting from the combination of OCTA and OCT B-scan based major voting models showed a testing accuracy of 94.0%.

Inception-ResNet-v2 allows the best prediction of early exudation of qMNV in AMD. Ensembling techniques further improve the accuracy of the classification, resulting in a model with >90% accuracy in prediction when combining OCT B-scan and OCTA images.

Artificial intelligence shows high performances in identifications of qMNVs at risk for exudation within the first 2 years of follow up. Better results are obtained with the combination of OCTA and OCT B-scan image analysis. The application of this model in clinical practice would allow better customization of follow up timing and would help avoiding treatment delay.