Radiomic Characterization and Automated Classification of Drusen Substructure Phenotype Associated with High-Risk Dry Age-Related Macular Degeneration

TL;DR

This study introduces automated radiomic metrics to classify drusen substructures in dry age-related macular degeneration, improving prediction of geographic atrophy risk.

Contribution

The novel use of radiomic features for automated drusen classification and GA risk prediction is introduced.

Findings

Radiomic features classified drusen phenotypes with AUC = 0.87–0.95.

H-type drusen show higher reflectivity and coarser texture compared to others.

Radiomic features predict GA conversion and growth rate with AUC = 0.59–0.74.

Abstract

Background/Objectives: Optical coherence tomography (OCT)-reflective drusen substructures (ODSs) are associated with the conversion of intermediate AMD to geographic atrophy (GA). However, ODSs must be manually identified, a laborious process introducing bias and variation. This study proposes objective radiomic metrics of drusen phenotypes and validates them for the prediction of GA development and GA growth rate. Methods: A total of 104 drusen with high-reflective cores (H-type), 105 with low-reflective cores (L-type), 129 conical drusen (C-type), and 101 normal drusen (N-type) were segmented from OCT images. Radiomic features were extracted from these drusen, and the most important features for drusen classification were extracted from the retinal pigment epithelium–Bruch’s membrane compartment of 743 OCT scans of eyes with dry AMD and used to predict GA conversion and fast growth. Results: Radiomic features classified drusen phenotypes with AUC = 0.87–0.95. H-type drusen have a higher reflectivity, greater variation in reflectivity, and coarser texture (p < 0.001). L-type drusen have a lower reflectivity and greater variation in reflectivity (p < 0.0001). C-type drusen have a less spherical shape and more disordered internal reflectivity (p < 0.001). N-type drusen have a more spherical shape and more uniform internal reflectivity (p < 0.001). These radiomic features predict the conversion from intermediate AMD to GA and top-quartile GA growth rate with AUC = 0.59–0.74 at years 1–3. Conclusions: These results demonstrate the potential of clinical phenotype-grounded radiomics for objective automated drusen analysis, GA risk stratification, and clinical prediction.

Figures