A deep learning pipeline for the detection and quantification of drusen and reticular pseudodrusen on optical coherence tomography

Abstract

Purpose : In recent years, reticular pseudodrusen (RPD) have been identified as risk factors for advanced age-related macular degeneration (AMD). Only two studies explored machine learning (ML) techniques for their automatic detection on OCT, neither allowing for accurate lesion quantification. We developed a deep learning pipeline for the detection and quantification of conventional drusen (CD) and RPD in the UK Biobank (UKBB), a large-scale biomedical database and research resource.

Methods : A ML pipeline consisting of several ML models was developed: a. A classification model and an out of distribution model for ungradable scans. b. A classification model to identify scans with drusen of any type (CD or RPD). d. An image segmentation model to independently segment lesions as RPD or CD, allowing their quantification.

Of 2,622 UKBB participants with a self-reported diagnosis of AMD, 1,284 had OCT scans and were included in the study. Manual delineation of features was performed by five experienced graders for the segmentation model. We used 3D Inception-V1 as the architecture for the classification models and a 2D U-Net architecture for the segmentation model.

To measure overlap in segmented areas between the model and graders we used free-response receiver operating characteristic (FROC) curves. In addition, the intraclass correlation coefficient (ICC) for absolute agreement was used to measure agreement in the area of the different lesions between the model and graders and for interrater reliability analysis.

Results : The first classification model classified ungradable scans with an area under the curve (AUC) of 94%. The deep ensemble models achieved an AUC of 87%. The second classification model achieved an AUC of 99.3% for classifying drusen of any type vs. controls. Considering one false positive lesion per image on average, the segmentation model achieved a sensitivity of 85% for drusen and 62% for RPD. The ICC for CD area was 0.79 and for RPD it was 0.66. This exceeded human interrater agreement (0.72 for CD, 0.37 for RPD).

Conclusions : We present a deep learning pipeline for the elimination of ungradable scans, classification of drusen, and segmentation of CD and RPD. To the best of our knowledge, this is the first pipeline enabling the total workflow, and this is the first DL model to allow accurate quantification of these lesions exceeding human performance.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.