Radiomics-based fundus autofluorescence analysis in central serous chorioretinopathy-MICRoN report number twelve.

This study applies radiomics-based feature extraction to fundus autofluorescence (FAF) images to automatically identify quantitative biomarkers that distinguish simple versus complex and acute versus chronic subtypes of central serous chorioretinopathy (CSCR). A total of 96 FAF images representing different CSCR stages were analyzed: simple (n = 47), complex (n = 49), acute (n = 47), and chronic (n = 49). Fifty-two radiomic features were extracted from the entire 55° FAF images using Pyfeats. Variables with zero variance and highly correlated features (Pearson's r > 0.8) were excluded. The most informative features were identified and used to construct and evaluate logistic regression, random forest, and extreme gradient boosting (XGBoost) classifiers for CSCR stage prediction. For simple versus complex CSCR, eight selected features yielded excellent discriminative performance in the logistic regression model, with a mean area under the receiver operating characteristic curve (AUC) of 0.90 (95% CI, 0.71-1.00) and an accuracy of 0.80. Sensitivity and specificity were balanced at 0.80 each, indicating robust and stable classification performance. For acute versus chronic CSCR, ten features were selected. The XGBoost model demonstrated modest discriminative ability with a mean AUC of 0.69 (95% CI, 0.39-0.98) and an accuracy of 0.71, with balanced sensitivity (0.70) and specificity (0.71) across resamples. Radiomic features extracted from FAF images effectively distinguish simple and complex CSCR, supporting their potential as quantitative imaging biomarkers for automated CSCR classification and disease staging.