Slab selection for effective en-face OCT segmentation of hypertransmission defects (hyperTD) in neovascular age-related macular degeneration (nAMD)

Abstract

Purpose : Finding suitable slab depths and thicknesses for hyperTD analysis on en-face optical coherence tomography (OCT) for human grading and computer vision in nAMD eyes.

Methods : We used 15 97-line OCT volumes (Heidelberg Spectralis) from 15 nAMD eyes, creating 30 en-face OCTs per volume with varying slab depths (0-300 µm below Bruch’s Membrane (BM)) and thicknesses (25-336 µm).
To identify the most suitable slabs for human grading of hyperTD, an ophthalmologist selected 3 slabs per volume that best met 4 criteria: high contrast, clear hyperTD delineation, minimal choroidal vessel interference, and low artifact presence (Fig 1A).
HyperTD was manually annotated as ground truth (GT) on the en-face image from the reference slab (64 to 400 µm below BM) (Fig 1B), and was classified into 3 distinct phenotypes with clear, less defined borders and subtle structural changes.
To identify suitable slabs for automated computer vision segmentation, we tested 3 algorithms (Thresholding (TH), Chan-Vese (CV) and Watershed (WS)) on different slabs, assessing their performance against the manual GT (Fig 1C).

Results : HS showed slabs 200-250 µm below BM best met the 4 criteria (Fig 2A). Automated segmentation was less slab-location-dependent (Fig 2B).
For hyperTD with distinct borders, best WS algorithm segmentation was at 100-300 µm below BM, with a DS of 0.44. HS preferred a narrower slab of 200-225 µm.
For hyperTD with less distinct borders, the CV algorithm performed better at the 100-300 µm slab below BM with a DS of 0.65 aligning with HS.
For subtle-structure hyperTD, CV was most effective at a 200 to 225 µm slab below BM, with a DS of 0.36. HS for this phenotype suggested a thicker slab (336 µm) 25 µm below BM, for better visibility.

Conclusions : For manual annotations, a 200 - 250 um slab below BM may generate the best en-face OCT for hyperTD analysis. Automated segmentation was generally less slab-dependent. In summary, a thin, deep slab is suitable for hyperTD annotation and segmentation in nAMD.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.

 

Fig 1 - A) En-face of focal hyperTD with varied slabs and blue-boxed human selections (HS), B) Manual annotation from the slab with red box, C) WS segmentation with green boxes for highest Dice scores (DS).

 

Fig 2 - A) HS of the slabs, B) median DS for different algorithms over different slabs