W e thank Dr Wang for the thoughtful comments on our article and for raising the important issues of reproducibility, standardization, and clinical scalability in quantitative optical coherence tomography angiography (OCTA) and ultra-widefield fluorescein angiography (UWF-FA) analysis. We agree that these issues are central to the future translation of retinal imaging biomarkers. However, we respectfully disagree with the implication that automation alone ensures superior reproducibility or validity compared with carefully standardized manual or semiautomated assessment.
First, several of the methodological concerns raised in the Correspondence were already addressed in the Methods section of our article and in a prior publication on which our analytic pipeline was based. , OCTA image averaging was used as a standardized postprocessing approach to improve signal-to-noise ratio and reduce speckle noise, rather than as image manipulation intended to alter biologic findings. As described previously, the superficial capillary plexus angiogram with the highest signal intensity and minimal motion artifact was selected as the reference. A rigid model was used for feature extraction, followed by elastic registration to align the superficial capillary plexus slabs. The same transformation was then applied to the deep capillary plexus slab before averaging. Thus, this workflow was designed to reduce motion-related and speckle artifacts rather than introduce additional analytic bias.
Similarly, the calculation of geometric perfusion deficits was based on previously described methods and biologically motivated definitions. , In our workflow, averaged angiograms were binarized using the Huang thresholding method, capillary maps were skeletonized, and geometric perfusion deficits were defined as retinal tissue located >30 μm from the nearest perfused capillary. Large-vessel masks and the foveal avascular zone were excluded from the relevant measurements to reduce distortion from projection, shadowing, and physiological nonvascular areas.
We agree that automation is an important future direction. Nevertheless, automated segmentation algorithms require rigorous validation before they can be assumed to outperform manual or semiautomated methods. This is particularly relevant in diabetic retinopathy, where irregular capillary dropout, low signal, media opacity, and disrupted retinal morphology can compromise automated algorithms. Importantly, automated segmentation and annotation systems are typically developed, trained, and validated against expert manual annotations. In this sense, expert manual annotation remains the ground truth framework on which many automated approaches are built and evaluated. Manual assessment should therefore not be viewed simply as an inferior alternative to automation but rather as a necessary reference standard, particularly in research settings requiring careful phenotyping.
With respect to UWF-FA quantification, manual delineation of retinal nonperfusion remains a standard approach in contemporary retinal vascular imaging studies. Large multicenter reading-center studies, including the DRCR Retina Network Protocol AA, have determined retinal nonperfusion and gradable area manually using ImageJ by masked graders, with nonperfusion index calculated from these measurements. Thus, the use of manual tracing in our study reflects the current state of rigorous UWF-FA research rather than a methodological departure from standard practice.
Overall, we agree that greater automation and multicenter validation will be important for future translation. At the same time, we believe that carefully standardized manual and semiautomated methods remain appropriate for research-grade quantification of OCTA and UWF-FA biomarkers in diabetic retinopathy.
See the original article for any disclosures of the authors.
CRediT authorship contribution statement
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
