Abstract
This paper proposes an automated, real-time method to detect a patient’s position inside an MRI scanner using low-resolution "move during scan" data. By accurately identifying body regions like the head, neck, and thorax without manual input, the system can optimize Specific Absorption Rate (SAR) limits, leading to safer, faster, and higher-quality imaging.
Introduction
Magnetic Resonance Imaging (MRI) is non-invasive, but the radio frequency (RF) pulses used to create images generate heat in the patient’s tissues, measured as the Specific Absorption Rate (SAR). To prevent injury, international standards strictly limit SAR levels based on the body part being imaged; for instance, the head has significantly different thermal tolerances than the torso.
Currently, MRI systems rely on radiology staff to manually enter patient data—weight, height, and orientation—and physically align the patient to a specific spot on the bed. This process is tedious, prone to human error, and often results in "conservative" safety margins. If the scanner cannot verify the patient's exact position, it defaults to lower power settings to ensure safety, which unfortunately reduces image quality or increases scan time. This research aims to replace this manual bottleneck with an automated image-processing solution.
Methods and Materials
The authors developed a slice-based detection pipeline designed to run in real-time as the patient bed moves into the bore. They utilized a "move during scan" protocol, providing low-resolution MR data (64x64 pixels) sufficient for anatomical localization but computationally inexpensive.
Anatomical Localization
- Thorax and Neck Detection: Since lungs are filled with air, they appear as dark cavities in MR images. The algorithm uses thresholding and region growing to identify these cavities. By calculating the area of these "lung holes" relative to the total slice area, the system identifies the thorax. The neck is localized by finding the transition zone (gradient descent) between the head and the thorax.
- Orientation Classification: To distinguish between "head-first" and "feet-first" orientations, the authors used Principal Component Analysis (PCA) for dimensionality reduction. Slices were downsampled to a 16x16 grid and projected into a 50-dimensional "eigen-space."
- Mahalanobis Distance: New slices are compared against training sets for "head" and "feet" classes in the reduced PCA space. Decisions are made by accumulating distance scores (a "trust value") across multiple slices.
Dataset
The study utilized whole-body scans from 13 test subjects. Four datasets served as training sets for the PCA and Mahalanobis classifiers, while nine were used for validation. Subjects were positioned with various offsets and supports (e.g., pillows) to test robustness. Data was acquired with a resolution of 7.5mm (x, y) and 7.5mm to 15mm (z).
Results and Discussion
The PCA-based classifier successfully detected orientation for 100% of the test subjects. The lung-cavity method for thorax detection worked accurately in 11 out of 13 cases. The primary failures were attributed to clothing artifacts, such as metal buttons or zippers, which caused signal dropouts that disrupted the lung cavity topology.
Key Insights
- The Neck as a Critical Transition: Identifying the neck accurately allows the scanner to safely increase power for the torso while maintaining strict limits for the brain.
- Efficiency of Low-Resolution Data: By proving that 16x16 pixel downsampling is sufficient for safety-critical decisions, the authors demonstrated that the algorithm can run on existing hardware without adding time to the clinical workflow.
- Inherent Robustness: Unlike optical cameras, MR-based detection sees internal anatomy, making it less susceptible to interference from blankets or clothing (excluding metallic artifacts).
Conclusion
The proposed automated system offers a robust path toward removing manual data entry in MRI setup. Future work will focus on improving robustness against metallic clothing artifacts, potentially through Naive Bayes classifiers, and expanding detection to include specific joints like the knees or pelvis.
Summary updated