# Radiomics-Based OCT Analysis of Choroid Reveals Biomarkers of Central Serous Chorioretinopathy

## TL;DR

This study uses radiomics to analyze OCT images of the choroid and identifies biomarkers for central serous chorioretinopathy, improving automated diagnosis.

## Contribution

The study introduces a radiomics-based method for automatically identifying choroidal biomarkers in central serous chorioretinopathy using OCT imaging.

## Key findings

- Radiomics features achieved 84.2% accuracy in distinguishing healthy from CSCR eyes using horizontal OCT images.
- En face images showed 85.3% classification accuracy for healthy versus CSCR eyes.
- Radiomics features revealed CSCR signatures in unaffected fellow eyes with high classification accuracy.

## Abstract

Biomarkers from choroidal imaging can enhance clinical decision-making for chorioretinal disease; however, identification of biomarkers is labor-intensive and limited by human intuition. Here we apply radiomics feature extraction to choroid imaging from swept-source optical coherence tomography (SS-OCT) to automatically identify biomarkers that distinguish healthy, central serous chorioretinopathy (CSCR), and unaffected fellow eyes.

Radiomics features were extracted from SS-OCT images from healthy (n = 30), CSCR (n = 39), and unaffected fellow eyes (n = 20), with a total of 44,500 single-cross sectional horizontal images and 8900 en face images. Logistic regression classification of eyes as healthy versus CSCR, healthy versus fellow, or CSCR versus fellow was performed using radiomics features. Statistical significance was determined using 95% bootstrap confidence intervals.

Significant differences between healthy and CSCR eyes were found for all radiomics feature groups. Classification of health versus CSCR achieved classification accuracy of 84.2% (77.2%–89.9%) in horizontal images and 85.3% (78.2%–90.7%) in en face images. For en face images, classification accuracy increased by 1.02% (0.50%–1.53%) for every 10% increase in choroid depth. Fellow eye classification using a classifier trained to distinguish healthy and CSCR eyes resulted in 90.4% (90.2%–90.6%) of horizontal images and 90.2% (89.8%–90.2%) of en face images being classified as CSCR.

These results demonstrate accurate classification of healthy and CSCR eyes using choroid OCT radiomics features. Furthermore, radiomics features revealed signatures of CSCR in unaffected fellow eyes.

These findings demonstrate the potential for radiomics features in clinical decision support for CSCR.