Automated choroidal neovascularization associated abnormality detection and quantitative analysis from clinical SD-OCT

Abstract

 

Purpose

 

To evaluate a fully-automated fluid-associated abnormality detection (SEAD) method on a large SD-OCT dataset. The volume of choroidal neovascularization associated abnormalities has the potential to better guide anti-VEGF therapy in CNV. This is an evaluation of a previously published technique on a large cohort of patients.

 

Methods

 

Subjects with choroidal neovascularization (CNV) due to AMD were imaged the same day and macular scans (Zeiss Cirrus HD-OCT) with 200×200×1024 voxels covering 6×6×2 mm³ were obtained. Ten retinal layers are segmented using the Iowa Reference Algorithm, followed by SEAD footprint detection yielding projections of fluid-filled regions. A supervised classification method follows to generate initial SEAD probability in 3D. A combined graph search/graph cut method is used to segment pairs of SEAD-adjacent retinal layers and any present SEAD region in 3D. The independent standard was generated by an ophthalmologist who manually marked all voxels inside SEADs. The performance was evaluated by a 3D overlap method yielding Dice coefficient values. Pearson’s correlation was used to measure the degree of linear correlation between the quantified SEAD volumes given by the automated method and the independent standard.

 

Results

 

Seventy SD-OCT scans from 23 subjects were evaluated. Our method succeeded to detect SEADs in 57 of the 70 scans. For 12/57 scans that exhibited small SEAD volumes (<0.18 mm³), a modest value of the Dice coefficient was obtained (0.44). For larger SEADs in 45/57 scans (volume > 0.18 mm³), the Dice coefficient was 0.76, implying a good overlap agreement between the automated and manual segmentations (see also Figs. 1, 2). The Pearson’s correlation for large SEADs was 0.95.

 

Conclusions

 

A state-of-the-art automated fluid-associated abnormality segmentation method was evaluated. Our method demonstrated good SEAD detection ability across SEAD sizes and good detection and quantitative assessment performance for larger-sized SEADs. Automated quantification has the potential to guide the treatment of patients with anti-VEGF agents and lead to optimized frequency of repeat injections.

 

 

Figure 1: Visualization of SEAD segmentation. A) Original OCT, B) Independent standard, and C) Automated segmentation (the segmentation is inherently 3D).

 

 

Figure 2: Performance assessment: automated SEAD segmentation (y-axis) and independent standard (x-axis).

 

Keywords: 549 image processing • 550 imaging/image analysis: clinical • 412 age-related macular degeneration