Development and validation of a multi-modal MRI-based deep learning framework for differentiation of intraspinal tumors (ISMF-Net).

Background: Intraspinal tumors (ISTs) pose diagnostic challenges due to their complex anatomy and reliance on subjective clinical imaging interpretation, while artificial intelligence-based methods offer significant potential for non-invasive ISTs diagnosis. We aimed to develop and validate an MRI-based deep learning model for IST differentiation. Methods: We conducted a retrospective study across three hospitals in China between January 2010 and July 2025. This retrospective study included 1004 patients diagnosed with ISTs, categorized into Schwannoma (SCN), Meningioma (MNG), Astrocytoma (AST), Ependymoma (EPN), and Metastasis (MET). Patients from Center 1 (n = 723) formed the internal dataset, which was randomly divided into training, validation and internal test sets at a 7:1:2 ratio. Additionally, 281 patients from Centers 2 and 3 were used as an external test set. All patients underwent preoperative spinal MRI, including sagittal T1-weighted (T1W), T2-weighted (T2W), T2 fat-suppressed (T2FS) and axial T2W sequences. We proposed the ISMF-Net that integrates features from sequences and clinical data to differentiate ISTs. Diagnostic improvements with model assistance were evaluated through an observer study involving radiologists of different levels of experience. Findings: Among 1004 patients, 22.6% had SCN, 21.8% had MNG, 14.9% had AST, 18.2% had EPN, and 22.5% had MET. The proposed ISMF-Net achieved micro-ACC of 0.859 [0.827, 0.889], 0.849 [0.826, 0.874], and 0.821 [0.801, 0.841] on the validation, internal test, and external test sets, respectively, outperforming existing methods. The highest performance was observed for SCN (F1: 0.905 [0.879, 0.930]-0.940 [0.910, 0.968]), while MET showed lower performance (F1: 0.757 [0.724, 0.790]-0.804 [0.746, 0.858]). Model assistance significantly improved diagnostic performance across all radiologists, with junior radiologists benefiting the most, showing increases of 7.9%, 20.1% and 4.9% in accuracy, sensitivity, and specificity, respectively. Additionally, preference analysis confirmed strong alignment between the model's predictions and radiologists' diagnostic tendencies. Interpretation: This study proposes a multi-modal DL approach integrating MRI and clinical data to improve IST diagnosis, providing a valuable tool for enhancing diagnostic accuracy in clinical practice. Funding: This study was funded by the General Program from Department of Education of Liaoning Province (JYTMS20230132), the National Natural Science Foundation of China (82304235) and the China Postdoctoral Science Foundation (2024M753641).