Comparison of synthetic CT generation algorithms for MRI-only radiation planning in the pelvic region
2018
Accurate radiation dose calculation is a major challenge in magnetic resonance imaging (MRI)-only radiation therapy (RT) treatment planning as the required electron density map is not readily provided by this modality. In this work, a number of state-of-the-art synthetic-CT (sCT) generation methods, exhibited promising results in the literature, were evaluated based on common quantitative metrics and patients dataset. This includes four atlas-based approaches, specifically median of atlas images (A-Median) [1], atlas-based voxel-wise weighting (A-VW) [2], bone enhanced atlas-based voxel-wise weighting (A-Bone) [3], iterative atlas-based voxel-wise weighting (A-Iter) [4], and a method based on deep learning convolutional neural network (DL-CNN) [5]. Automatic organ delineation was performed for bladder, rectum and bone. Overall, A-VW, A-Bone, A-Iter and A-VWexhibited comparable performance while DL-CNN showed slightly better segmentation performance resulting in Dice metrics of 0.93, 0.90, and 0.93, respectively. The dosimetric evaluation demonstrated that A-Median, A-VW, A-Bone, A-Iter and DL-CNN resulted in comparable mean dose errors within organs at risk and target volumes showing less than 1% dose difference against the CT-based RT planning. The two-dimensional gamma analysis performed at 1%/1 mm criterion demonstrated comparable pass rates of 94.99±5.15%, 94.59±5.65%, 93.68±5.53% and 93.10±5.99% for A-Bone, DL-CNN, A-Median and A-Iter, respectively. Whereas A-VW and water-only resulted in pass rates of 86.91±13.50% and 80.77±12.10%, respectively. DL-CNN and advanced atlas-based approaches showed promising dosimetric and segmentation accuracy (DL-CNN is slightly better) suggesting that these methods are able to resolve the challenge of synthetic-CT generation from MR images with clinically acceptable errors.
Keywords:
- Correction
- Source
- Cite
- Save
- Machine Reading By IdeaReader
9
References
3
Citations
NaN
KQI