Novel semi-automated system for multi-dimensional analysis of macular neovascularization: quantitative biomarkers and...

To develop and validate a novel semi-automated Python-based pipeline for multi-dimensional analysis of macular neovascularization (MNV) on optical coherence tomography angiography (OCTA) images. The system provides standardized biomarkers, rule-based morphological categorization, and morphology-derived interpretive labels. It also incorporates an automated morphological classification scheme to objectively distinguish active, mature quiescent, transitional, and arteriolarized states.

We developed an advanced image-processing pipeline incorporating hybrid multiscale vessel enhancement using Frangi/tubeness and Laplacian of Gaussian filters, hybrid binarization (LoG-Otsu and tubeness-Sauvola with small-particle removal), dynamic region-of-interest refinement, and boundary-branch exclusion. The system generates a Standardized Vascular Complexity Score, a Standardized Caliber Uniformity Score, and a Standardized Maturity Index on a 0-100 scale, and performs rule-based classification into Medusa, Seafan, Glomerular, Tree in bud, and Dead tree patterns using predefined percentile and trunk-pattern thresholds. Morphology-derived interpretive categories - Active pattern, Maturequiescentpattern, Transitional-pattern, and Arteriolarized pattern - were assigned using quantitative thresholds and published morphological criteria. A total of 112 MNV lesions acquired on three OCTA platforms (Zeiss PlexElite 6 × 6 mm, Zeiss CIRRUS AngioPlex HD 3 × 3 mm, and Optovue Solix 6 × 6 mm) were analyzed. Masked expert-automated morphological agreement was assessed in a stratified subset (n = 54). Inter-observer reproducibility was evaluated with three operators in 46 lesions using ICC(2,1), and same-operator intra-observer test-retest reproducibility was assessed as a supplement (n = 46). Between-stratum score differences were summarized using Kruskal-Wallis tests and ε² effect sizes with bootstrap 95% confidence intervals.

Raw topological metrics differed substantially by device, precluding direct numerical comparison. Device-specific standardization using principal component analysis and piecewise-linear normalization yielded convergent standardized Complexity, Caliber Uniformity, and Maturity scores on a common 0-100 reporting scale across devices, with median values of approximately 50. Under this standardized framework, Complexity Score, Caliber Uniformity Score, and Maturity Index did not differ significantly across strata (Kruskal-Wallis p = 0.425, 0.572, and 0.582, respectively; ε² ≈ 0). Automated morphological classification identified the full spectrum of MNV subtypes across all platforms. Agreement with masked expert grading was modest to moderate, with an overall agreement of 57.4% and a quadratic weighted κ of 0.507. Inter-observer ICC(2,1) was good for lesion area (0.859), Complexity Score (0.807), and Standardized Caliber Uniformity Score (0.770), and moderate for the Maturity Index derived from the Caliber score (0.593).

This system provides a transparent and robust platform for quantification and rule-based morphological reporting of MNV architecture across OCTA platforms, establishing a common framework for comparative analysis. Future outcome-linked and multi-center evaluations will be required to establish clinical utility. By integrating morphological patterns with quantitative biomarkers and automated classification, the system enables objective differentiation of active angiogenesis from mature remodeling.