Optical coherence tomography (OCT) is a non-invasive imaging technology widely used in ophthalmic diagnostics, particularly for retinal and optic nerve evaluation (Huang et al., 1991). Compared to traditional imaging, OCT offers superior resolution and speed, enabling real-time visualization of microstructural changes for early disease detection (Costa et al., 2006). Beyond ophthalmology, OCT has shown potential in systemic diseases such as Alzheimer’s and Parkinson’s (Gaire et al., 2024; Ge et al., 2021; Wagner et al., 2023).

Advancements in OCT technology have led to spectral-domain OCT (SD-OCT) and swept-source OCT (SS-OCT), which differ in scanning speed, penetration depth, and segmentation algorithms (Laíns et al., 2021). These differences result in significant inter-device variations in key metrics like central subfield thickness (CST) (Nam et al., 2024; Matlach et al., 2014), retinal nerve fiber layer (RNFL) (Matlach et al., 2014; Pierro et al., 2012; Seibold et al., 2010), ganglion cell inner plexiform layer (GCIPL) (Mahmoudinezhad et al., 2021), limiting cross-study data integration (Tan et al., 2016). Standardized conversion equations are thus critical for harmonizing multi-device datasets (Bressler et al., 2014; Lee et al., 2021; Schrems et al., 2015; Shepherd et al., 2012).

Artificial intelligence (AI) has emerged as a powerful tool in ophthalmology (Foo et al., 2021; Wang et al., 2023; Shahriari et al., 2023; Feng et al., 2023; Li et al., 2024; Xu et al., 2022; Xu et al., 2020). Methods based on machine learning or deep learning can learn patterns or regularities from data through algorithms, achieving robust predictive or decision-making performance. At present, Cirrus HD-OCT 5000 and Triton DRI-OCT are the SD-OCT and SS-OCT, respectively, commonly used in clinical practice, and the studies on the measurement differences between the two machines are mostly limited to a certain amount of measurement indexes and do not give a clear method of data standardization (Nam et al., 2024; Lee et al., 2017). Therefore, this study is the first to jointly evaluate the differences between SD-OCT and SS-OCT for whole-zone measurements of macular retinal thickness, RNFL, and GCIPL, and to construct machine learning-based standardized equations to address data compatibility issues in clinical studies and multicenter collaborations.

A total of 96 eyes of 48 healthy adults (including 13 males and 35 females) with a mean age of 32.9 ± 13.95 years were included in this study.

In this study, we found that there were large differences in the measurements between Triton DRI-OCT and Cirrus HD-OCT 5000 between macular retinal thickness in the macula, macular GC-IPL thickness, and RNFL thickness, which is consistent with previous studies (Nam et al., 2024; Lee et al., 2017). There are multiple reasons for such measurement differences, including differences in SS-OCT and SD-OCT scanning parameters, image segmentation algorithms, and measurement areas (Laíns et al., 2021; Pierro et al., 2012; Heussen et al., 2012), such as in the measurement of the CST, the Cirrus HD-OCT 5000 when measuring the distance from the ILM to the intermediate layer of the RPE, and the Triton DRI-OCT when measuring the distance from the ILM to the OS and the RPE boundaries, and when measuring RNFL thickness, the Cirrus HD-OCT 5000 measured the average RNFL thickness in a 3.46 mm diameter circular area centered on the optic disk, while the Triton DRI-OCT measured a 3.4 mm diameter circular area (Lee et al., 2017).

We also compared the consistency of the measurements between the two devices, which did not have consistent ICC indices between each measurement, with the lowest ICC index located in the thickness of the nasal sector of the RNFL (0.451, P < 0.001) and the highest ICC index in the thickness of the outer subcircular region of the macular retina and the nasal region (0.906, P < 0.001). In contrast, Nam et al. (2024) found that the Triton OCT-1 and Cirrus HD-OCT 5000 had a higher ICC index in the temporal region of the outer retinal ring in the macula (0.930, 95% CI 0.828–0.972), which may be related to the population differences and segmentation algorithms between the machines. Overall, both devices showed high agreement in measured retinal thickness, RNFL thickness, and GCIPL thickness in the macular region.

Nonetheless, the significant differences between the two devices on the same measurements resulted in their measurements not being directly merged for use. Therefore, standardized conversion of data in clinical studies involving multiple OCT devices is mandatory (Shepherd et al., 2012; Kenney et al., 2022). Based on four machine learning methods, we constructed a model for standardized conversion of OCT measurement data, in which the model showed the best predictive performance for CST, with R² ranging from 0.892 to 0.930 in the testing set, and the worst predictive performance for RNFL, with R² ranging from 0.516 to 0.760 in the testing set, which may be mainly due to the difference between the two machines in the RNFL and the GCIPL with different segmentation criteria (Lee et al., 2017), which resulted in the measures showing large differences. Moreover, it is noteworthy that the study cohort exclusively comprised healthy adults, whose clear and regular retinal images do not currently challenge OCT segmentation algorithms. However, in patients with ocular pathologies such as diabetic macular edema (DME) or wet age-related macular degeneration (wAMD), irregular retinal structures frequently induce OCT segmentation errors. This may lead to significant discrepancies in measurements between different devices. Consequently, researchers have initiated efforts to incorporate diseased populations—specifically establishing data conversion equations for CST in DME patients (Bressler et al., 2014; Bressler et al., 2015; Ibrahim et al., 2012), and developing conversion equations for macular retinal thickness in AMD patients (Krebs et al., 2011). Therefore, some researchers have pinned their hopes on reducing the measurement differences of multiple OCT devices by constructing a standardized automatic segmentation model based on Cycle GAN for image segmentation and layering of multiple OCT devices (Chen et al., 2023), but this research is still affected by the inconsistent image resolution and scanning artifacts of different OCT devices. Building upon the above discussion, while the data conversion equations derived from healthy retinae in this study require further validation for extension to patients with fundus diseases, this conversion model holds substantial potential for research on retinal alterations associated with extraocular pathologies such as Alzheimer’s disease (AD), mild cognitive impairment (MCI), multiple sclerosis (MS), and cerebrovascular disorders.

Among these four machine learning models, this study observes that LR is more stable in its performance on CST, RNFL, and GCIPL in both training and testing sets, and the R² in the testing set is 0.930, 0.845, and 0.760. This finding aligns with previous research outcomes. Schrems et al. (2015), established conversion equations for RNFL thickness between SD-OCT and TD-OCT using multiple methodologies, including non-conversion, mean difference, mean ratio, and linear regression approaches. Among these, linear regression demonstrated optimal predictive performance. Although numerous prior studies have endeavored to develop OCT data conversion equations(9, 29), few have employed machine learning paradigms such as nonlinear modeling. Our results similarly indicate that nonlinear models confer no significant advantage in constructing data conversion equations. Therefore, constructing the data conversion equation between two OCT machines using a linear model is a simple and effective standardization method. On the other hand, previous studies on OCT data conversion have been largely confined to comparisons between SD-OCT and TD-OCT. Kenney et al. (2022), through a generalized linear estimation model proposed data transformation equations for Cirrus and Spectralis SD-OCT on pRNFL and GCIPL, i.e., Cirrus = −5.0 + 1.0 × Spectralis pRNFL and Cirrus = −4.5 + 0.9 × Spectralis GCIPL; Bressler et al. (2014) constructed data conversion equations for the CST of patients with diabetic macular edema measured using TD-OCT and SD-OCT, which were found to further reduce measurement error. Sun et al. (2021), on the other hand, used a linear mixed model to construct the CST of patients with diabetic macular edema between the Cirrus and Spectralis SD-OCT transformation equation, i.e., Spectralis = 40.78 + 0.95 × Cirrus, and good predictive performance was also observed. This study concentrates on measurement discrepancies between SS-OCT and SD-OCT. Following the development of data conversion equations, empirical validation demonstrates their capacity to substantially mitigate these inter-device variations.

This study posits that data conversion equations can address two critical issues in clinical research. Firstly, the integration of OCT data in multicenter studies: Data conversion equations enable standardized transformation of raw measurements from different OCT devices, allowing merged analysis or comparison of standardized multicenter data to minimize measurement discrepancies. Secondly, OCT data standardization in longitudinal follow-up studies: Equipment replacement during extended follow-up introduces significant measurement errors that compromise data integrity; applying these conversion equations permits standardization of historical data in select studies, thereby mitigating inter-device variability. While maintaining consistent measurement protocols remains the gold standard for robust clinical research, these conversion equations must be judiciously applied when unavoidable technical transitions occur.

Nonetheless, the limitations of this study are as follows: only healthy adults were included in this study, and differences between race, age (Tham et al., 2020) and disease (Bressler et al., 2014) were not considered, and the applicability of the constructed data conversion model to patients with macular or optic nerve disease still requires further investigation; in addition, the conversion equations for data measured by different machines are different (Bressler et al., 2014; Lee et al., 2021; Schrems et al., 2015), and the data conversion model constructed in this study is limited to the Triton DRI-OCT and Cirrus HD-OCT 5000 machines; furthermore, the data conversion model constructed to combine the measurement data from different OCT devices still has overfitting or underfitting, which requires caution in actual clinical research, but there are certain differences in scanning parameters, image segmentation algorithms, and measurement areas between the devices. However, there are some differences in scanning parameters, image segmentation algorithms and measurement areas between devices, and it is still helpful to construct a data conversion model to minimize the measurement error in clinical research.

Our study found that there are significant differences in macular retinal thickness, GCIPL, and RNFL thickness between the Triton DRI-OCT and the Cirrus HD-OCT 5000, and the construction of a machine learning model can effectively reduce this measurement error, which provides a reliable data standardization method for cross-sectional studies or long-term follow-up studies using different OCT devices.