Automatic Calcified Drusen Segmentation in Swept-Source and Spectral-Domain OCT Scans using Deep-Learning

Abstract

Purpose : Calcified drusen assessments are clinically important for determining the risk of age-related macular degeneration (AMD) progression. This study aims to develop an automated deep-learning-based algorithm to segment calcified drusen, applicable to both swept-source (SS) and spectral-domain (SD) optical coherence tomography (OCT) scans.

Methods : Eyes exhibiting calcified drusen underwent imaging with SS-OCT and/or SD-OCT utilizing 6x6mm scanning patterns. Customized composite en face images were generated using optical attenuation coefficient (OAC) converted OCT images. Three deep-learning models, employing U-Net architecture, were trained (Fig.1). The first model was trained from composite images generated from the drusen map and OAC mean projection between retinal pigment epithelium (RPE) and Bruch's membrane (BM). The second was trained by composite images generated by the drusen map, OAC mean projection between RPE and BM, and OAC sum projection calculated from internal limiting membrane (ILM) to BM. The third was trained by composite images generated by the drusen map, OAC mean projection between RPE and BM, and OCT mean projection obtained from 64 to 400 μm below BM (known as sub-RPE slab). Manual masks of calcified drusen were generated by identifying choroidal hypotransmission defects (hypoTDs) on en face structural images using subRPE slabs and confirmed by the presence of hyperreflective contents within drusen on corresponding B-scans. Model performances were assessed using DICE similarity coefficients (DSCs).

Results : In total, 101 SS-OCT and 82 SD-OCT volume scans were included in this study. In the context of joint training with both SS-OCT and SD-OCT, Model 3 demonstrated highest DSCs, achieving 76.81% in the testing dataset, as expected because human grading was based on sub-RPE slab. However, Model 2 provided a higher DSC (60.90%) compared to that of Model 1 (51.35%). Examples are shown in Fig.2

Conclusions : The proposed deep-learning model, using OAC information together with the calcified drusen appearance on sub-RPE slab, delivers the best performance in segmenting the calcified drusen in both SS-OCT and SD-OCT scans. However, a model trained from the OAC information alone without considering choroidal information delivered satisfactory results, which is clinically useful, especially for SD-OCT scans where light penetration into choroid is limited.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Fig.1

 

Fig 2