Comparative performance of radiomics models based on 3D-FLAIR-CUBE and MPRAGE sequences gadolinium-induced labyrinth MRI protocols for predicting endolymphatic hydrops in Ménière's disease.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Minde Road No. 1, Nanchang, 330006, Jiangxi Province, China.
- Jiangxi Provincial Key Laboratory of Intelligent Medical Imaging, Nanchang, 330006, Jiangxi Province, China.
- Department of Otolaryngology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Minde Road No. 1, Nanchang, 330006, Jiangxi Province, China.
- Department of Radiology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Minde Road No. 1, Nanchang, 330006, Jiangxi Province, China. [email protected].
- Jiangxi Provincial Key Laboratory of Intelligent Medical Imaging, Nanchang, 330006, Jiangxi Province, China. [email protected].
Abstract
This study aimed to develop and validate a radiomics-based machine learning model using 3D-FLAIR-CUBE and 3D MPRAGE gadolinium-induced labyrinth MRI protocols and accurately predict endolymphatic hydrops in unilateral Meniere's disease. This retrospective study enrolled 175 patients with definite unilateral MD (2019-2023) who underwent three-dimensional cube liquid attenuation inversion recovery (3D-FLAIR-CUBE) and three-dimensional magnetization-prepared rapid gradient echo (3D-MPRAGE) MRI after intratympanic gadolinium injection. Radiomic features (n = 1666) from cochlear and vestibular regions were extracted using PyRadiomics, followed by feature selection the least absolute shrinkage and selection operator, (LASSO) regression and Pearson correlation. Three machine learning models logistic regression (LR), Naive Bayes (NB), and multilayer perceptron (MLP) were trained (80% data set) and tested (20%) to classify EH. Model performance was evaluated by AUC, accuracy, calibration curves and decision curve analysis (DCA). In the modeling based on 3D-FLAIR-CUBE radiomic features, the machine learning models demonstrated strong capability in identifying the affected ear. The multilayer perceptron (MLP) model achieved an AUC of 0.914 in the training set and 0.815 in the test set. Analysis based on 3D-MPRAGE features yielded slightly lower yet still significant performance, with AUCs of 0.778 in the training set and 0.712 in the validation set. A combined model integrating features from both sequences showed AUCs of 0.913 and 0.814 for the training and test sets, respectively. However, DeLong's test indicated that the combined model did not provide a statistically significant improvement compared to the model using the CUBE sequence alone. Notably, the model based on 3D-FLAIR-CUBE outperformed that based on 3D-MPRAGE. Decision curve analysis further confirmed its favorable clinical net benefit. Importantly, feature selection revealed that vestibular features carried greater diagnostic weight than cochlear features in the final model, suggesting their higher relevance in identifying endolymphatic hydrops. This study developed the first radiomics model based on gadolinium-induced labyrinth MRI using 3D-FLAIR-CUBE and 3D-MPRAGE sequences. The MLP model based on 3D-FLAIR-CUBE demonstrated the best diagnostic performance. Focusing on vestibular imaging features can improve the detection of endolymphatic hydrops. This tool may assist in guiding personalized treatment decisions, particularly for patients with refractory Ménière's disease.