En Face OCT Detection and Segmentation of Drusen Using Multi-Stage Deep Learning Algorithms

Abstract

Purpose : We propose a multi-stage deep learning pipeline for automated drusen detection and segmentation on en face OCT imaging at the Bruch’s membrane (BM) layer.

Methods : A retrospective dataset of en face swept-source OCT from healthy subjects (20) and patients with dry AMD (22) were analyzed. Exclusion criteria included geographic atrophy, exudative lesions, and high myopia. Volumetric 12 x 12-mm scans centered on the macula were acquired, and BM segmentation was performed using an established automated algorithm with volumetric smoothing at the post-processing stage. 50 adjacent en face slices spanning 0 to 100 microns above the BM were generated by flattening normalized and shadow compensated B-scans at the BM layer. Final en face images were created by averaging the generated slices, cropping at the 6 x 6-mm center, and resizing to 256 x 256 pixels. Ground truth masks of drusen were obtained by manual segmentation and verified by an expert clinician. The deep learning pipeline included a ConvNeXt model in the first stage for classification of the presence or absence of drusen followed by ResUNet and Pix2Pix generative adversarial network (GAN) models in the second stage for segmentation of drusen. An ensemble segmentation model was evaluated using the union of the generated segmentations. A random 80:20 training-testing data split was used with data augmentation performed on the training set. The performance of the models was assessed using accuracy for classification and dice coefficient (DC) for segmentation.

Results : 50 OCT volumes, 27 with drusen and 23 healthy, were studied. Training data was expanded to 4000 en face images after data augmentation. Training and testing accuracies for classification were 99.76% and 100.00% respectively. Mean testing DC values (± standard deviation) for the
ResUNet, Pix2Pix GAN, and the ensemble models were 0.6933 (± 0.0692), 0.6992 (± 0.0692), and 0.7023 (± 0.0756) respectively.

Conclusions : Deep learning pipelines leveraging GAN architectures yield accurate drusen identification and segmentation on en face OCT despite variability in ground truth data. Automated evaluation of drusen subtypes is underway.

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

 

Proposed deep learning pipeline for classification (green) and segmentation (blue) of drusen in en face OCT images.

 

Proposed drusen segmentation in en face OCT using ResUNet and Pix2Pix GAN architectures as well as an ensemble model.