Work

Standardized conversion model for retinal thickness measurements between spectral-domain and swept-source optical coherence tomography based on machine learning

Year: 2025

Type: article

Abstract: Purpose To conduct a systematic comparative analysis of macular retinal thickness, retinal nerve fiber layer (RNFL) thickness, and ganglion cell-inner plexiform layer (GCIPL) thickness measurements between spectral-domain optical coherence tomography (SD-OCT) and swept-source OCT (SS-OCT) in healthy individuals, while establishing standardized cross-platform conversion algorithms through machine learning methodologies. Methods In this cross-sectional investigation, 48 healthy adults (96 eyes) underwent macular retinal thickness assessment (ETDRS grid sectors), RNFL analysis (quadrant sectors), and GCIPL evaluation (six-sector annular divisions) using both SD-OCT (Cirrus HD-OCT 5000) and SS-OCT (Triton DRI-OCT). Inter-device measurement differences were evaluated through paired t-tests. Agreement metrics were quantified via intraclass correlation coefficients (ICCs) and Bland-Altman analysis. Four predictive models—linear regression (LR), LASSO regression, random forest regression (RF), and support vector regression (SVR)—were developed to estimate Triton DRI-OCT measurements from Cirrus HD-OCT 5000 outputs. Model efficacy was assessed using coefficient of determination (R2) and root mean square error (RMSE). Results Statistically significant inter-device discrepancies (P < 0.001) were identified in 9 macular sectors, all GCIPL parameters (average and six-sector measurements), and RNFL measurements (average thickness and three quadrants, excluding nasal sector). ICC values demonstrated moderate-to-strong agreement: macular thickness (0.771–0.906), GCIPL (0.554–0.710), and RNFL (0.451–0.852). Machine learning models exhibited superior performance in central subfield thickness (CST) prediction, achieving test set R2 values of 0.930 (LR), 0.926 (LASSO), 0.936 (SVR), and 0.892 (RF). Linear regression maintained consistent predictive accuracy across parameters: CST (R2 = 0.930), RNFL (R2 = 0.845), and GCIPL (R2 = 0.760). Conclusion Substantial measurement discrepancies preclude direct interchangeability of SD-OCT and SS-OCT datasets. Machine learning-derived conversion algorithms significantly improve cross-device comparability, offering a robust standardization framework for multicenter research and longitudinal data integration. This methodological advancement enables harmonized analysis of OCT metrics across heterogeneous imaging platforms.

Source: Frontiers in Cell and Developmental Biology

DOI: 10.3389/fcell.2025.1612455