Abstract
Purpose :
Machine learning models based on radiomic feature extraction from clinical imaging data provide effective and interpretable means for clinical decision-making, yet have had limited application to ophthalmic imaging. This study evaluated whether radiomics features in baseline optical coherence tomography (OCT) images of eyes with pigment epithelial detachment (PED) associated with wet age-related macular degeneration (wAMD) can predict treatment response to as needed anti-VEGF (vascular endothelial growth factor) therapy.
Methods :
OCT images were obtained from 25 eyes with PED at baseline, month 3, and month 6 during as-needed anti-VEGF therapy. Radiomics features were extracted using the Pyfeats python-based radiomics tool. PED response to treatment was defined by projecting image features onto an axis defined by the mean feature values for baseline and 3-month follow-up images. Eyes were labeled as unresponsive, regressing, or responsive based on projected feature values. Naive Bayes was used to classify baseline images as responsive at 6 months or not responsive at 6 months (i.e., unresponsive or regressing). To ask whether regressing eyes were more similar at baseline to unresponsive or responsive eyes, a second classifier was trained to classify eyes as responsive or unresponsive and applied to regressing baseline images. Classification performance was obtained using leave-one-out cross-validation, and statistical significance was assessed with an approximate permutation test (1000 iterations).
Results :
Radiomics feature analysis of 25 eyes identified 12 unresponsive eyes, 6 regressing eyes, and 7 responsive eyes. Naive Bayes classification of baseline features as responsive versus unresponsive or regressing resulted in classification performance of 84.0% (p <0.001). Classification of regressing eyes as responsive or unresponsive resulted in all regressing eyes classified as unresponsive.
Conclusions :
Our results demonstrate the use of radiomics features to identify eyes that are likely to respond to as-needed anti-VEGF therapy. Eyes regressing with as-needed treatment had similar baseline features to unresponsive eyes. Our study demonstrates the potential of radiomics features for effective and interpretable augmentation of clinical decision-making.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.