Fully-Automated Drusen Segmentation in OCT using Deep Learning with Pyramid U-net

Abstract

Purpose : The clinical hallmark of early and intermediate age-related macular degeneration (AMD) is the presence of drusen. Currently, the scheduling frequency of patients with dry AMD is primarily guided by the amount of drusen, which is assessed from images that are acquired noninvasively with the help of optical coherence tomography (OCT). Furthermore, quantitative drusen measurements such as drusen volume are becoming an important factor in assessing disease progression. In this work, we propose a new deep learning model for drusen segmentation in OCT.

Methods : We developed a U-net convolutional neural network (CNN) to segment outer retinal layers and drusen. Spatial pyramid pooling was introduced into the CNN to improve the automated segmentation model. To enable training and validation, 435 OCT volume scans of 38 patients (50 eyes) with intermediate AMD (AREDS II or higher) acquired with Heidelberg Spectralis were included. The dataset reference standard was created by manual annotations of retinal pigment epithelium (RPE) and Bruch’s Membrane (BM) by experienced ophthalmologists (Fig. 1). B-scans were supplied to the CNN without any preprocessing and outputs were evaluated without post-processing.

Results : The segmentation network was trained and evaluated using ‘leave-one-patient-out’ cross-validation. Qualitative evaluation showed that the proposed deep learning model was able to segment drusen in a precise and reproducible manner (Fig. 2). We quantitatively evaluated the segmentation performance by computing the Dice score measuring the overlap between the segmentations and the annotated reference standard. The evaluation showed that the U-net CNN achieved a mean patient Dice score of 0.75.

Conclusions : The proposed deep learning method is an important step towards an accurate quantification of drusen, which is crucial for the successful image-based clinical management of patients with early/intermediate AMD. There is an urgent need for advanced medical image computing methods that can measure distinct and pathognomonic changes of drusen morphology in an objective and reproducible manner to assess the conversion risk to later AMD stages and prognosis determination.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

OCT B-scan intensity image (left), and its reference annotation (right) with the three classes: Drusen (blue), RPE (red) and BM (yellow).

 

Two examples of the drusen segmentation result with the BM in green and the lower RPE boundary in red.