A custom algorithm can be used to identify inflammatory choriocapillaris lesions with a high degree of agreement with human graders, and outperformed these graders in reproducibility, according to research results published in the American Journal of Ophthalmology.
Researchers conducted an observational case series to develop and validate an automated algorithm that could be used to independently identify and segment the boundary of clinically relevant flow deficit lesions in patients with posterior uveitis. Additional aims were to compare the performance of the algorithm with 2 human graders.
Included patients had a diagnosis of posterior uveitis with choroidal involvement who underwent swept-source optical coherence tomography angiography (SS OCT-A) between 2016 and 2020. Two trained human graders performed manual choriocapillaris lesion boundary identification; to evaluate intra-grader repeatability testing, graders were also asked to draw lesion boundaries on the same images across 3 separate sessions each 24 hours apart. The algorithm was also used to perform boundary identification 3 times.
Images included 20 eyes from 14 participants (71% men) with choroidal inflammatory diseases including birdshot chorioretinopathy (n=4), serpiginous choroiditis (n=4), multifocal choroiditis (n=3), acute multifocal posterior pigment epitheliopathy (n=1), relentless placoid choroiditis (n=1), and presumed ocular histoplasmosis (n=1).
Lesions from 20 eyes with posterior uveitis were segmented by the algorithm and 2 masked human graders to test the ability of the algorithm to perform segmentation. For each image, total lesion area was identified, and the spatial overlap of the lesions were compared.
Overall, the algorithm demonstrated “excellent” agreement with the human graders in terms of determining total lesion area (ICC, 0.92; 95% CI, 0.81-0.97; ICC 0.91; 95% CI, 0.78-0.97 for human grader 1 and 2, respectively).
Average lesion area from all eyes measured by the algorithm, though, was significantly smaller than the average lesion area measured by the algorithm (7.30 mm2 vs 10.72 mm2 and 10.60 mm2, respectively).
The algorithm also demonstrated good spatial overlap with at least 1 human grader in 80% of eyes, and with both human graders in 70% of eyes.
Poor spatial overlap on the part of the algorithm was seen in 20% of eyes vs the human graders. These eyes were qualitatively evaluated for possible common features that might explain this outcome. Two of the eyes contained a large choroidal neovascular membrane, of which all or most was excluded by the algorithm.
Researchers also sought to determine if the algorithm performed with a similar reliability as the human graders. The algorithm demonstrated perfect intra-grader repeatability in terms of delineating the same lesion boundary on the same image, while neither human grader demonstrated perfect repeatability.
A set of unaltered images from outside of the study institution were captured and analyzed in order to demonstrate the algorithm’s generalizability and utility. In 4 patients, the algorithm was able to segment all lesions and provide a single quantitative measurement of each eye’s disease burden.
Study limitations include the susceptibility of OCTA imaging to motion artifacts and speckle noise, a lack of “ground truth” for choriocapillaris imaging, and the use of a curated dataset of high-quality images that might differ from the images available in clinical practice.
“Automated lesion localization and area quantification were highly reliable and reproducible in comparison with human graders,” the investigators explain. “Real-time, objective, and quantitative monitoring of choroidal disease activity may be possible using this automated image analysis approach.”
Disclosure: Some study authors declared affiliations with the biotech, pharmaceutical, or device companies. Please see the original reference for a full list of authors’ disclosures.
McKay KM, Chu Z, Kim J-B, et al. Automated quantification of choriocapillaris lesion area in patients with posterior uveitis. Am J Ophthalmol. Published online June 6, 2021. doi:10.1016/j.ajo.2021.06.004