{
  "abstract": "Background Accurate differentiation of intramedullary spinal cord (IMSC) pathologies remains challenging on routine MRI, often necessitating invasive diagnostic procedures with non-negligible morbidity. Deep learning (DL) offers a potential non-invasive solution by extracting latent imaging features beyond conventional radiological interpretation.Objective To develop and validate an MRI-based DL framework capable of differentiating three major IMSC entities: spinal dural arteriovenous fistula (DAVF), multiple sclerosis (MS), and spinal cord tumors, using routine clinical MRI sequences.Methods 77 adult patients (29 MS, 23 DAVF, 25 tumors) who underwent spinal MRI between 2007 and 2024 were retrospectively included. For each case, contrast-enhanced T1-weighted (T1W+C) and T2-weighted (T2W) images were analyzed, with manual spinal cord segmentation performed by a senior neurosurgeon. Multiple convolutional neural network architectures (ResNet-18, ResNet-50, EfficientNet-B0) were trained using transfer learning. An extensive ablation study evaluated nine MRI input configurations, including multimodal combinations and anatomically guided masking. Five-fold cross-validation was applied. Model performance was assessed using accuracy, F1-score, and ROC-AUC.Results ResNet-18 demonstrated the most robust and consistent performance across input configurations. The optimal input combination, T2W images combined with spinal cord mask and intensity-mean channel, achieved an F1-score of 0.87 ± 0.04 and ROC-AUC of 0.95 ± 0.03. Class-specific performance was highest for DAVF (AUC = 0.99), followed by tumors (0.96) and MS (0.90). Confusion matrix analysis showed high sensitivity across all classes, with particularly strong discrimination of vascular lesions.Conclusions An MRI-based deep learning approach can reliably differentiate major intramedullary spinal cord pathologies using routine imaging alone. This non-invasive framework demonstrates strong diagnostic performance and has the potential to reduce reliance on diagnostic angiography or biopsy, accelerate diagnosis, and add further support for the clinical decision-making in the complex management of spinal cord lesions.Disclosures O. Haim: None. S. Gaby: None. M. Arzi: None.Abstract E-351 Figure 1Abstract E-351 Figure 2",
  "authors": [
    {
      "affiliations": [
        "Neurosurgery, University of Utah, Salt Lake City, UT",
        "Neurosurgery, Tel Aviv University, Tel Aviv, ISRAEL"
      ],
      "name": "O Haim"
    },
    {
      "affiliations": [
        "Neurosurgery, Tel Aviv University, Tel Aviv, ISRAEL"
      ],
      "name": "S Gaby"
    },
    {
      "affiliations": [
        "Sagol Brain Institute, Tel Aviv University, Tel Aviv, ISRAEL"
      ],
      "name": "M Arzi"
    }
  ],
  "title": "E-351 MRI-based deep learning for prediction of spinal cord lesions",
  "uid": "1d0a04eb-0c84-5346-bdc0-246bdb0a51c4"
}