Multi-view and multimodal radiological grading using spinal MRIs

Conference item

This paper proposes a transformer-based model that encodes MRI volumes from multiple sequences and anatomical views of the spine, and predicts multiple spinal gradings. The transformer ingests slice-wise 2D embeddings and a learnable class token to capture the relationships between the slice-wise embeddings. The method is applied to predict finegrained radiological gradings of spinal stenosis conditions (spinal canal stenosis, right and left neural foraminal narrowing, and right and left subarticular stenosis) using T1-weighted, T2-weighted and STIR sequences in sagittal and axial views. Experiments show that our joint multi-view, multimodal model outperforms task-specific baselines trained on individual modalities or views across all tasks.

Authors: Park, RY; Windsor, R; Jamaludin, A; Zisserman, A

• Publication status: Published
• Peer review status: Peer reviewed
• Acceptance date: 2025-08-16
• Event title: 28th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2025)
• Event location: Daejeon, South Korea
• Event website: https://conferences.miccai.org/2025/en/default.asp
• Event start date: 2025-09-23
• Event end date: 2025-09-27
• Language: English
• Keywords: multimodal learning; radiological grading; stenosis