Feasibility of AI-Based Wheelchair Classification for MRI Anteroom Safety Management: A Single-Centre Deep Learning Object-Detection Study.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiological Technology, Nagoya City University East Medical Center, 1-2-23 Wakamizu, Chikusa-Ku, Nagoya, 464-8547, Japan. [email protected].
- Department of Radiological Technology, Nagoya City University East Medical Center, 1-2-23 Wakamizu, Chikusa-Ku, Nagoya, 464-8547, Japan.
Abstract
In MRI examinations, MR-unsafe metallic wheelchairs inadvertently brought into the scanner room represent a serious safety hazard. This study was aimed at developing and evaluating an automated wheelchair detection and classification system for the MRI anteroom using a lightweight deep learning object-detection model (YOLOv8n). Images were collected over 5 working days from two surveillance cameras in the MRI anteroom of a single academic medical centre. A dataset of 1,730 images (training: 1251; validation: 118; test: 361) was constructed from frames containing MR-safe and MR-unsafe wheelchairs, with 25-pixel mosaic anonymisation applied for privacy. Models were trained at 640 px and 1024 px resolutions with three random seeds each and evaluated on the independent test set. The 640 px models achieved [email protected]:0.95 of 0.702 ± 0.012 and real-time CoreML inference of 41.4 FPS on Apple M4 hardware, numerically outperforming 1024 px models across most metrics (the 95% confidence interval for the between-resolution difference in Precision included zero). Critical misclassification rates (MR-unsafe misidentified as MR-safe) reached 8.3% in the worst-case model when evaluated against MR-unsafe instances only (n = 72), which represents the clinically relevant denominator, and ranged from 0.21% to 1.28% (mean 0.64%) when expressed relative to all annotated (ground-truth) instances (n = 468). End-to-end sensitivity for MR-unsafe wheelchairs, which accounts for both missed detections and class errors, was 81.5% (range 76.4-86.1%). This single-centre proof-of-concept study demonstrated the technical feasibility of deep learning-based wheelchair classification for MRI anteroom safety screening and characterised its failure-mode profile. The findings derive from one anteroom layout, one camera configuration, and 72 MR-unsafe instances, and are hypothesis-generating rather than deployment-ready; external multi-centre validation is a prerequisite for clinical use.