Weilin Yang ^1,2 Xiaorui Su ³ Shuang Li ¹ Kaiyang Zhao ⁴ Qiang Yue ^1*

• ¹ Department of Radiology, West China Hospital, Sichuan University, 37# Guo Xue Xiang, Chengdu, Sichuan 610041, PR China., Chengdu, Sichuan, China
• ² Department of Radiology, Key Laboratory of Birth Defects and Related Diseases of Women and Children, Ministry of Education, West China Second University Hospital, Sichuan University, Chengdu, Sichuan, China
• ³ Department of Radiology, West China Hospital of Medicine, Huaxi MR Research Center (HMRRC), Chengdu, Sichuan 610041, PR China., Chengdu, Sichuan, China
• ⁴ West China Hospital of Sichuan University, #37 Guo Xue Xiang, Chengdu Sichuan, Chengdu, Sichuan 610041, PR China., Chengdu Sichuan, China

Objective: To develop a machine learning-based clinical and/or radiomics model for predicting the primary site of brain metastases using multiparametric magnetic resonance imaging (MRI). Materials and Methods: A total of 202 patients (87 males, 115 females) with 439 brain metastases were retrospectively included, divided into training sets (brain metastases of lung cancer [BMLC] n=194, brain metastases of breast cancer [BMBC] n=108, brain metastases of gastrointestinal tumor [BMGiT] n=48) and test sets (BMLC n=50, BMBC n=27, BMGiT n=12). A total of 3404 quantitative image features were obtained through semi-automatic segmentation from MRI images (T1WI, T2WI, FLAIR, and T1-CE) . Intra-class correlation coefficient (ICC) was used to examine segmentation stability between two radiologists. Radiomics features were selected using analysis of variance (ANOVA), recursive feature elimination (RFE), and Kruskal–Wallis test. Three machine learning classifiers were used to build the radiomics model, which was validated using five-fold cross-validation on the training set. A comprehensive model combining radiomics and clinical features was established, and the diagnostic performance was compared by area under the curve (AUC) and evaluated in an independent test set. Results: The radiomics model differentiated BMGiT from BMLC (13 features, AUC = 0.915 ± 0.071) or BMBC (20 features, AUC = 0.954 ± 0.064) with high accuracy, while the classification between BMLC and BMBC was unsatisfactory (11 features, AUC = 0.729 ± 0.114). However, the combined model incorporating radiomics and clinical features improved the predictive performance, with AUC values of 0.965 for BMLC vs. BMBC, 0.991 for BMLC vs. BMGiT, and 0.935 for BMBC vs. BMGiT. Conclusion: The machine learning-based radiomics model demonstrates significant potential in distinguishing the primary sites of brain metastases, and may assist screening of primary tumor when brain metastasis is suspected whereas history of primary tumor is absent.