Program Information
Dual Compartment Regression for the Generation of Pseudo-CT Images
AK Schnurr2, E Orasanu1*, H Schulz1, H Nickisch1, S Renisch1, (1) Philips Research Laboratories, Philips GmbH Innovative Technologies, Hamburg, Germany, (2) Otto-von-Guericke-University Magdeburg, Magdeburg, Germany,
Presentations
SU-K-601-4 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: 601
Purpose: To assess the feasibility of generating pseudo CT images of the pelvis from magnetic resonance imaging using a two-compartment regression approach.
Methods: Using a model-based segmentation of the bone structures in mDixon MR data sets of the pelvis separating the bone from the soft tissue compartment, Gaussian Process Regression is used to learn and subsequently predict Hounsfield units based on features (Soft tissue: patches, Bones: patches and geometry information) from mDixon image sets. The regression is trained on rigidly registered pairs of CT and MR data, which are affinely registered to an atlas. Samples are selected from the training images using a histogram-based scheme.For validation the voxelwise Mean Absolute Error (MAE) between pseudo CT and the patient CT was computed in a Leave-One-Out-Cross-Validation on 17 data sets. In addition, prostate EBRT plans which were available on the original CTs were re-simulated on the pseudo CTs, and the differences were analyzed.
Results: The MAE inside the body contour was 61.5266±109.3748HU. For bone structures in the pelvis a MAE of was 200.1777±205.2707 achieved. The method reliably differentiates between fat and soft tissue, resulting in a MAE of 46.791±80.801 for non-bone voxels. Excluding voxels in the patient CT, which have intensities outside of the range of fat and soft tissue [-200HU…200HU], for the comparison results in a MAE of 39.0129±32.4678.Re-simulated EBRT plans using the pseudo CT were evaluated using Gamma analysis. Out of the voxels receiving more than 75% of the maximum dose, on average 99.8% (98.6%...100%) have a gamma value <1 using a 1%/1mm gamma criterion.
Conclusion: This study indicates that a combination of bone/soft tissue separation and a Gaussian Process Regression per compartment is a promising approach to improve on currently available methods using bulk density assignment or single regression.
Funding Support, Disclosures, and Conflict of Interest: All authors are employed by Philips GmbH Innovative Technologies, which is a subsidiary of Royal Philips NV.
Contact Email: