KEY POINTS
- Phantom-based repeatability and reproducibility study of magnetic resonance imaging-derived radiomics on a 0.35 tesla magnetic resonance linear accelerator.
- The study used a tissue-mimicking phantom with 4 material types: fat-equivalent oil, external water, 40% polyvinylpyrrolidone, and fibroglandular dense-tissue mimic.
- Investigators acquired 50 balanced steady-state free precession scans across 10 sessions and extracted 68 radiomic features from 12 predefined regions of interest.
- 48 preprocessing pipelines were tested, combining normalization, discretization, and optional N4 bias-field correction.
- N4 bias correction plus z-score normalization with scaling around 128-256 produced the highest overall stability; homogeneous water and polyvinylpyrrolidone inserts were more reproducible than fat and fibroglandular compartments.
CLINICAL TAKEAWAY
This study reinforces a basic but often ignored point: magnetic resonance imaging-derived radiomics is highly preprocessing-dependent, even in a controlled phantom setting. For magnetic resonance linear accelerator-based biomarker work, N4 bias correction plus z-score normalization appears technically attractive, but these results do not validate any clinical biomarker. Clinical translation still requires patient-level validation across scanners, institutions, anatomy, and treatment timepoints.