Tong Wang ^1,2* Hongliang Mao ³ Yi Wang ³ Yi Wang ³

• ¹ Anhui University, Hefei, China
• ² School of Physics and Optoelectronic Engineering, Anhui University, Hefei, Anhui, China, Hefei, China
• ³ Department of Neurosurgery, First Affiliated Hospital of Anhui Medical University, Hefei, Anhui Province, China

Trigeminal neuralgia (TN) is among the most severe neuropathic disorders, necessitating precise diagnosis for personalized treatment. Imaging studies typically lack clear objective markers. Computer-assisted diagnosis provides a non-invasive, quantitative alternative, although existing methods face limitations in availability, accuracy, and reliability. This study employs a multimodal fusion approach to enhance diagnostic precision, particularly in identifying idiopathic trigeminal neuralgia. We retrospectively collected MRI data and clinical information from 342 patients with trigeminal neuralgia and normal control (NC) from two independent centers. The MRI modalities included T2-weighted and 3D Time-of-Flight (TOF) sequences. Features were extracted from the MRI data by using radiomics and deep learning techniques, and feature selection was performed using LASSO regression. In order to build a new method for diagnosing TN with high accuracy and high reliability, we extracted two modalities features from different MRI imaging modalities. we built our prediction model in one dataset and do external validation on another independent dataset. We subsequently constructed a clinical diagnostic model based on a single imaging modality using Support Vector Machine. Finally, based on the disease probabilities predicted by these unimodal models, we developed a post-fusion model using Stochastic Gradient Descent methodology. The results from the external validation cohort indicate that the Cli + Rad in post-fusion models outperforms the singlemodal and pre-fusion models in predictive performance. (Accuracy 0.821; AUC = 0.837; Sensitivity = 0.813; recall = 0.864 and specificity = 0.826.). In this study, we proposed a new method for prediction. The results demonstrate significant improvements in all measured parameters of the post-fusion model, providing a scientific basis for the personalized treatment of TN and enabling the standardization of the diagnostic process.