Based on the linked source; automatically prepared and checked against the original report.
Safety rationale
MR-unsafe metallic wheelchairs inadvertently brought into MRI scanner rooms represent a serious safety hazard. Researchers aimed to develop and evaluate automated detection and classification in the MRI anteroom using the lightweight YOLOv8n object-detection model. The study reports technical feasibility and should not be interpreted as a treatment recommendation.
Study design
Images were collected over five working days from two surveillance cameras in the MRI anteroom of one academic medical centre. The dataset comprised 1,730 images: 1,251 for training, 118 for validation, and 361 for testing. Images contained MR-safe and MR-unsafe wheelchairs and received 25-pixel mosaic anonymisation. Models were trained at 640-pixel and 1,024-pixel resolutions with three random seeds, then evaluated on an independent test set.
Performance and validation needs
The 640-pixel models achieved mAP@0.5:0.95 of 0.702 ± 0.012 and 41.4 frames-per-second CoreML inference on Apple M4 hardware, numerically outperforming 1,024-pixel models across most metrics. In the worst case, 8.3% of 72 MR-unsafe instances were misidentified as MR-safe. End-to-end sensitivity ranged from 76.4% to 86.1%. The authors call the results hypothesis-generating because they derive from one layout, camera configuration, and 72 MR-unsafe instances; external multi-centre validation is required before clinical use.
