Radiomics features from optical coherence tomography of choroid differ in healthy and central serous chorioretinopathy patients

Abstract

Purpose : Choroid biomarkers provide predictive information about a wide range of disease states. However, manual identification of such biomarkers is labor intensive and often subjective. Here we use automated radiomics feature analysis to identify quantitative features that distinguish between healthy and central serous chorioretinopathy (CSCR) eyes.

Methods : SS-OCT images were obtained from healthy (n = 25) and CSCR (n = 13) eyes. Images for each eye consisted of 1024 OCT B-scan slices through the macula. An automated segmentation algorithm was applied to identify pixels in OCT corresponding to choroid. Radiomics features were extracted from each slice using the Pyfeats python-based radiomics tool (52 total features). To compare individual features across eyes, average feature value was computed across slices for each eye and feature values were compared across disease states. Next the spatial properties of radiomics features were analyzed by computing the autocorrelation across slices within each individual. Autocorrelation area under the curve (AUC) was then computed and compared across disease states. All comparisons were made using a bootstrap confidence interval (confidence level 0.05, 1000 iterations, with Bonferroni correction for number of features).

Results : Comparison of 52 radiomics features across healthy and CSCR eyes revealed 24 features significantly different between healthy and CSCR eyes. Mean autocorrelation AUC across features was 291 [95% CI, 269, 315] in healthy and 389 [95% CI, 366, 408] in CSCR eyes, indicating greater feature correlations across space in CSCR compared to healthy eyes.

Conclusions : Our results demonstrate the utility of radiomic feature extraction for automatically identifying quantitative biomarkers that differ between healthy and CSCR states. Biomarker analysis revealed significant differences between healthy and CSCR eyes in terms of texture feature values and in terms of the spatial consistency of these features. Future work will focus on comparing features across disease subtypes and using features for prediction of disease state.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.