Real-World External Validation of Artificial Intelligence-Based Full-Vessel Segmentation for Intracoronary Optical Coherence Tomography

It externally validated a previously developed full-vessel segmentation algorithm in an unselected consecutive real-world cohort rather than only on well-curated datasets. Prior algorithms were mostly developed and validated on well-curated datasets; this study tested the algorithm on real-world data from 100 consecutive patients undergoing clinically indicated OCT, directly addressing the concern that such algorithms may not represent real-world data. Retrospective, single-center external validation design with 100 consecutive patients and 2560 analyzable frames, compared frame by frame against an expert manual reference standard.

Automated identification and quantification of calcified plaque reached an excellent level, close to interobserver variability. It provides concrete agreement metrics in real-world data for calcified plaque identification (κ=0.88 [95% CI, 0.84-0.92]) and quantification (intraclass correlation coefficients 0.79–0.93), explicitly noting performance close to interobserver variability. Frame-level agreement analysis across 2560 analyzable frames, reporting κ values and intraclass correlation coefficients with 95% confidence intervals.

Automated identification and quantification of lipid plaque was reasonable and largely superior to interobserver variability. It reports agreement for lipid plaque identification (κ=0.68 [95% CI, 0.64-0.72]), lipid arc (intraclass correlation coefficient 0.79 [95% CI, 0.76-0.81]), and minimum fibrous cap thickness (intraclass correlation coefficient 0.59 [95% CI, 0.55-0.63]), noting this was largely superior to interobserver variability. Also based on frame-level comparison, with κ values and intraclass correlation coefficients plus confidence intervals, benchmarked against interobserver variability.

Algorithm performance was limited for low-prevalence features such as plaque rupture. It explicitly identifies the limitation of automated identification for low-prevalence features in real-world data, delineating the current capability boundary for future work. The text states directly that 'algorithm performance for low-prevalence features (e.g., plaque rupture) was limited,' without giving specific values for that feature.

This is a single-center, retrospective external validation study of 100 consecutive patients undergoing clinically indicated OCT, suited to evaluating how the previously developed OCT-AID full-vessel segmentation algorithm performs on real-world consecutive-patient data; its findings support generalizability of the methodology and lay groundwork for subsequent validation across multiple centers, prospectively, and under different acquisition conditions.

The text does not give specific agreement values for low-prevalence features such as plaque rupture, nor does it report the distribution of each feature across the 2560 frames; in addition, performance across centers, devices, and prospective settings under a single-center retrospective design remains an open question to watch. Because the current reading scope is summary-level and does not include figures or supplementary material, these details cannot be further confirmed from the available text.