Evaluating an OCT reference database generated based on anatomical features

Abstract

Purpose : Normal anatomical variations in healthy eyes are often averaged out when reference databases (RDB) are created. In this study, we investigated the feasibility of generating an OCT RDB based on anatomical features and tested its use on healthy and glaucomatous eyes.

Methods : We obtained de-identified 12×9 mm widefield scans from Maestro2 (Topcon Corp., Toyko, Japan) devices at sites that screen patients with OCT. Vessel pattern features were extracted from each en face image. We used these as a sampling pool for two types of RDB for the RNFL thickness 2D map. One used 500 random scans from the sampling pool. The other RDB was anatomy-matched for each test scan by selecting 500 scans with the most similar features.
We constructed a set of test scans from two sources: glaucoma eyes from a tertiary clinic and normal eyes segregated from the sample set and reviewed using the Columbia University (CU) method (Prog Retin Eye Res 2022; 90:101052) to confirm normality. We then used the two RDB types to generate 2D deviation maps for each test scan. The number of pixels below the fifth percentile were computed and considered “flagged”.

Results : Over 7800 scans were used as the sampling pool. 29 healthy eyes and 31 glaucomatous eyes were used as test data. 83.9% (26/31) of glaucomatous eyes had increased flagging with anatomy-matched RDB, while the deviation maps for the 5 cases with less flagging still presented obvious glaucomatous damage. A student’s t-test showed significant differences (p = 6.55 x 10^-5) between the number of flagged pixels when using the two RDBs. For the healthy eyes, t-test showed no significant differences (p = 0.77). The average number of flagged pixels in this group accounted for 4.9% and 4.75% of the total pixels when using anatomy-matched and random RDB, respectively. These values are close to the 5% cutoff.

Conclusions : We demonstrated that anatomy related features such as vessel pattern can be extracted from OCT and be used to generate RDB. Current results suggest an anatomy-matched RDB has the potential to increase sensitivity for glaucoma detection while maintaining specificity. Further investigations are needed to test more data, especially eyes in early stage of glaucoma.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.

 

Three examples of glaucomatous eyes: (A) OCT en face images. (B) RNFL thickness maps overlaid with 30 μm isocontours from each RDB’s average RNFL thickness map. (C), (D) Deviation maps obtained using the two RDBs.