Radiomics-Based Machine Learning to Support Visual Assessment for Improved Epilepsy Classification Using 18F-FDG Brain PET

Abstract

Introduction: This study explored radiomics-based machine learning (ML) models as complementary tools to visual evaluation for classifying drug-resistant epilepsy patients and healthy controls using 18F-FDG brain Positron Emission Tomography (PET). Because visual interpretation can be subjective and variable, especially for novice readers, objective and reproducible computational methods are needed.Material and Methods: Twenty-one drug-resistant epilepsy patients and sixteen healthy controls underwent ¹⁸F-FDG brain PET imaging. From contralateral brain regions, 92 radiomics features (first-order statistics and second-order texture matrices) were extracted. Feature selection included Student’s t-test, principal component analysis, and ridge regression. Logistic regression (LR) and support vector machine (SVM) classifiers were trained and evaluated using 10-fold cross-validation and repeated 80/20 train–test splits. A permutation test (n = 1000) assessed whether differences between classifier performances were statistically significant. LR, chosen for its lower computational cost and interpretability, was used for comparison with human visual assessments.Results: Across six radiomics feature groups, LR models demonstrated strong performance, with mean accuracy of 0.94(0.05), precision 0.96(0.03), recall 0.92(0.10), specificity 0.97(0.02), and AUC 0.98(0.00). SVM models showed similarly high accuracy 0.98(0.01), precision 0.94(0.05), recall 0.96(0.03), specificity 0.98(0.01), and AUC 0.98(0.00). Novice visual assessments had moderate accuracy (0.62 and 0.67), perfect specificity, lower sensitivity (0.60 and 0.65), and AUCs of 0.80 and 0.825. The final LR model achieved a mean AUC of 0.96(0.01).Conclusion: This hybrid radiomics-visual approach improves classification accuracy in pre-surgical evaluation of drug-resistant epilepsy. By integrating quantitative radiomics with clinical interpretation, the framework reduces variability and improves reliability for less experienced clinicians.

Ähnliche Arbeiten