Stable and discriminating OCT-derived radiomics features for predicting anti-VEGF treatment response in diabetic...

Radiomics-based characterization of fluid and retinal tissue compartments of spectral-domain optical coherence tomography (SD-OCT) scans has shown promise to predict anti-VEGF therapy treatment response in diabetic macular edema (DME). Radiomics features are sensitive to different image acquisition parameters of OCT scanners such as axial resolution, A-scan rate, and voxel size; consequently, the predictive capability of the radiomics features might be impacted by inter-site and inter-scanner variations.

The main objective of this study was (1) to develop a more generalized classifier by identifying the OCT-derived texture-based radiomics features that are both stable (across multiple scanners) as well as discriminative of therapeutic response in DME and (2) to identify the relative stability of individual radiomic features that are associated with specific spatial compartments (e/g. fluid or tissue) within the eye.

A combination of 151 optimal responders and rebounders of anti-VEGF therapy in DME were included from the PERMEATE (imaged using Cirrus HD-OCT scanner) and VISTA clinical trials (imaged using Cirrus HD-OCT and Spectralis scanners). For each patient within the study, a set of 494 texture-based radiomics features were extracted from the fluid and the retinal tissue compartment of OCT images. The training set ( $S_t$ ) included 76 patients and the independent test set $(S_v$ ) comprised of 75 patients. Features were ranked based on (i) only discriminability criteria, that is, maximizing area under the receiver operating characteristic curve (AUC) and (ii) both stability and discriminability criteria. The subset of radiomic features for which the feature expression remained relatively consistent between the two datasets, as assessed by Wilcoxon rank-sum test, were considered to be stable. Different machine learning (ML) classifiers (such as k-nearest neighbors, Random Forest, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Support Vector Machine using linear and radial basis kernel, Naive Bayes) were trained using the features selected based on both the stability and discriminability criteria on $S_t$ and then subsequently validated on $S_v$ . The ML classifier ( $M_g$ ) that yielded maximum AUC on $S_v$ was considered to be more generalized and stable for distinguishing anti-VEGF therapy treatment response as well as less sensitive to the effect of inter-site and inter-scanner variability.

The model $M_g$ (based on both stability and discriminability criteria) achieved higher AUC compared to the criteria based off feature discrimination alone on $S_v$ (maximum AUCs of 0.9 versus 0.81; p-value = 0.048). The texture-based radiomic features pertaining to the retinal tissue compartment were found to be more stable compared to the fluid related features across the two datasets.

Our study suggests that incorporating both stable and discriminatory texture-based radiomic features extracted from fluid and retinal tissue compartments of OCT scans, a more generalized radiomic classifier can be developed to predict therapeutic response in DME. Also, the feature stability was found to be a function of the spatial location within the eye from where the features were extracted.