Two swept-source OCT machines disagreed on choroidal thickness by 65 microns

Design

device comparison study with vendor software and a common deep learning segmentation model, Bland-Altman and intraclass correlation analysis

Population

97 highly myopic Chinese adults imaged on two commercial swept-source OCT devices

Primary outcome

agreement in choroidal thickness across the ETDRS grid

Effect

vendor software global thickness 221.5 ± 66.9 against 156.9 ± 57.6 µm, mean differences 61-71 µm, intraclass correlation 0.63; with a common deep learning model, 173.5 against 171.2 µm, mean difference 2.26 µm, intraclass correlation 0.98

Choroidal thickness is increasingly used in high myopia, central serous chorioretinopathy and pachychoroid disease, and it is often quoted across studies and across clinics as though it were a single measurement. Ninety-seven highly myopic adults were imaged on two commercial swept-source OCT devices, with thickness measured across the ETDRS grid both by each vendor's own software and by a common deep learning segmentation model, with ocular magnification correction applied.

With vendor software, the two machines disagreed substantially: global thickness averaged 221.5 ± 66.9 µm on one device and 156.9 ± 57.6 µm on the other, with mean differences of 61 to 71 µm across regions (all P<0.001) and only moderate agreement (intraclass correlation 0.63, 95% CI 0.56-0.68). Running the same images through one deep learning model brought the two devices to 173.5 and 171.2 µm, a mean global difference of 2.26 µm with limits of agreement from -17.4 to +21.9 µm, and intraclass correlations above 0.90 in every sector and 0.98 globally.

A 65 µm systematic difference is larger than most of the changes anyone would act on. It means a choroidal thickness measured at one clinic cannot be compared with one measured at another, that a patient who changes provider appears to have a dramatically different choroid, and that published normative values are device-specific whether or not they say so. The finding also identifies where the error lives: it is the segmentation software, not the hardware, since a single model reconciled the two. Until vendors converge, the practical rule is that choroidal thickness is a within-device measurement only.

• Treat choroidal thickness as comparable only within the same device and software version
• Do not compare a patient's choroidal thickness against a published normative value derived on a different platform
• Where a patient transfers between clinics, re-establish a baseline rather than reading a change
• Record the segmentation method, since the disagreement lay in the software rather than the hardware
• Apply ocular magnification correction in high myopia, as this study did

Why it matters

A measurement being used clinically and in research turns out not to be the same measurement between machines.

Don't overread it

Two specific devices in 97 highly myopic Chinese adults - the size of the discrepancy is particular to these platforms and this population, though the principle is not.

The statistics, in plain English

An intraclass correlation of 0.63 means the two devices rank eyes similarly but assign different values - so a study using one machine can still find a valid association while its absolute numbers do not transfer. The Bland-Altman limits of agreement are the number to look at rather than the mean difference: even after the deep learning model brought the means within 2.26 µm, individual eyes still differed by up to about 20 µm, which is the realistic uncertainty on any single measurement. Ninety-seven highly myopic participants also means these figures apply to thin choroids; agreement may differ in normal eyes.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.