Hongjian Li ^1,2 Bing Li ¹ Chuan Zhang ¹ Ruhui Xiao ¹ Libing He ¹ Shaojie LI ³ Yu-Xin Yang ⁴ Shipei He ⁴ Baijintao Sun ¹ Zhiqiang Qiu ¹ Maojiang Yang ⁵ Yan Wei ¹ Xiaoxue Xu ¹ Hanfeng Yang ^1*

• ¹ Department of Radiology, Affiliated Hospital of North Sichuan Medical College, Nanchong, China
• ² North Sichuan medical college medical imaging college, Nanchong, China
• ³ Department of Neurosurgery, The Second Affiliated Hospital of Fujian Medical University Quanzhou, Fujian, China
• ⁴ Beijing United Imaging Intelligent Imaging Technology Research Institute, Beijing, Beijing Municipality, China
• ⁵ Department of Pain Medicine, Affiliated Hospital of North Sichuan Medical College, Nanchong, Shanxi Province, China

The diagnosis of primary trigeminal neuralgia (PTN) in radiology lacks the gold standard and largely depends on the identification of neurovascular compression (NVC) using magnetic resonance imaging (MRI) water imaging sequences. However, relying on this imaging sign alone often fails to accurately distinguish the symptomatic side of the nerve from asymptomatic nerves, and may even lead to incorrect diagnoses. Therefore, it is essential to develop a more effective diagnostic tool to aid radiologists in the diagnosis of TN.Purpose: This study aims to establish a radiomics-based machine learning model integrating multi-region of interest (multiple-ROI) MRI and anatomical data, to improve the accuracy in differentiating symptomatic from asymptomatic nerves in PTN.A retrospective analysis of MRI data and clinical anatomical data was conducted on 140 patients with clinically confirmed PTN. Symptomatic nerves of TN patients were defined as the positive group, while asymptomatic nerves served as the negative group. The ipsilateral Meckel's cavity (MC) was included in both groups. Through dimensionality reduction analysis, four radiomics features were selected from the MC and 24 radiomics features were selected from the trigeminal cisternal segment. Thirteen anatomical features relevant to TN were identified from the literature, and analyzed using univariate logistic regression and multivariate logistic regression. Four features were confirmed as independent risk factors for TN. Logistic regression (LR) models were constructed for radiomics model and clinical anatomy, and a combined model was developed by integrating the radiomics score (Rad-Score) with the clinical anatomy model. The models' performance was evaluated using receiver operating characteristic curve (ROC) curves, calibration curves, and decision curve analysis (DCA).The four independent clinical anatomical factors identified were: degree of neurovascular compression, site of neurovascular compression site, thickness of the trigeminal nerve root, and trigeminal pons angle (TPA). The final combined model, incorporating radiomics and clinical anatomy, achieved an area under the curve (AUC) of 0.91/0.90 (95% CI: 0.87-0.95/0.81-0.96) and an accuracy of approximately 82% in recognizing symptomatic and normal nerves.The combined radiomics and anatomical model provides superior recognition efficiency for the symptomatic nerves in PTN, offering valuable support for radiologists in diagnosing TN.