LapGM: A Novel Method to Accurately Bias Correct and Normalize MR Data
L Vinas1*, J Fischer2, A Sudhyadhom3, (1) University of California, Los Angeles, (2) University of Victoria, Calgary, AB, CA, (3) Brigham and Women's Hospital | Dana-Farber Cancer Institute | Harvard Medical School, Boston, MA
Presentations
SU-H300-IePD-F9-3 (Sunday, 7/10/2022) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]
Exhibit Hall | Forum 9
Purpose: To mitigate sharp bias fields generated from multi-coil MR scanners through a tunable Bayesian prior that allows for practitioners to optimize for visual clarity by balancing bias correction and image contrast. Intermediate byproducts of the method, such as estimated class labels, can be used as means to quantitatively compare MR intensity values across different patient scans. Additionally a multi-coil, arbitrary smoothness, bias generation procedure is implemented in closed form to more comprehensively evaluate bias correction methods.
Methods: A Bayesian bias correction method, LapGM, was developed and compared to N4ITK, the current standard for MR bias field correction, on simulated bias datasets and MRI pelvic patient scans. On simulated datasets, debiasing methods are compared on metrics: bias RMSE, intensity RMSE, tissue distributional total variation, and method runtime. On patient datasets, debiasing methods are compared qualitatively by comparing bias peak removal and contrast changes as no ground truth exists.
Results: LapGM surpasses N4ITK on all metrics in the multi-sequence setting. In the single-sequence setting LapGM performs comparably to N4ITK and can be simply tuned to marginally trade image contrast for targeted bias correction.
Conclusion: This work provides both a bias correction method and a procedure to generate multi-coil for arbitrary line geometries. In both cases interpretability was favored over complexity and it should be possible to improve performance considering advanced, adaptive updates and multi-resolution schemes. Both procedures are available in CPU and GPU accelerated formats for large image workloads.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email