Reproducibility of Optical Coherence Tomography Radiomics after Cataract Surgery: Effects of Media Opacity, Signal Strength, and Region of Interest Selection
- Author:
Na Hee KIM
1
;
Hyeong Min KIM
;
Hyewon CHUNG
;
Hyungwoo LEE
Author Information
- Publication Type:Original Article
- From:Korean Journal of Ophthalmology 2026;40(3):236-246
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:To determine how cataract-related image-quality improvement and region of interest (ROI) definition affect the reproducibility of optical coherence tomography (OCT) radiomics features.
Methods:This retrospective study included 50 eyes from 46 patients who underwent cataract surgery. Preoperative and postoperative OCT images were acquired using a standardized protocol. Radiomics features (n = 93) were extracted from two ROI configurations: retina ROI (Rr; retina only) and full ROI (Rf; retina, choroid, vitreous). Preoperative versus postoperative differences were tested with the Wilcoxon signed-rank test and Cohen’s d; repeatability was assessed with the intraclass correlation coefficient (ICC). Agreement between ROIs was evaluated with Wilcoxon statistics, Spearman correlation, the concordance correlation coefficient (CCC) and Bland-Altman plots. The influence of signal strength was analyzed with paired t-tests and Spearman correlation.
Results:Postoperative signal strength significantly improved from 28.24 ± 3.16 to 31.40 ± 3.06 dB (p < 0.001). Significant preoperative/ postoperative changes occurred in 49 / 93 features (52.7%) within Rr and in 81 / 93 (87.1%) within Rf. ICC showed good-to-excellent repeatability in 51.6% of Rr features and 87.1% of Rf features, yet 29.0% of Rr features still demonstrated poor reliability. Across ROI methods, 94.6% of feature pairs differed significantly between Rr and Rf, and no feature achieved CCC ≥0.85. Only two descriptors—first-order statistics 10th percentile and GLDM (gray level dependence matrix) gray level variance—showed no significant change with surgery and minimal dependence on signal strength.
Conclusions:Radiomics features in OCT respond variably to media opacity, signal-to-noise ratio, and ROI definition: most are susceptible to these factors, yet a small subset remains stable across them. Limiting analysis to an anatomically precise retina- bounded ROI and enrolling eyes with comparable cataract status and signal-strength scores may therefore improve reproducibility. Implementing these standardizations will help future OCT-radiomics studies identify the most reliable biomarkers for clinical use.
