Deep Learning-Based Synthetic CT from Black-Bone MRI for MR-Only Treatment Planning
Abstract
Purpose
Black-bone MRI provides improved contrast of the cortical bone than standard MRI sequences, and resembles a CT image when its contrast is inverted. Deep-learning (DL)-based synthetic-CT (sCT) techniques[1-2] often use standard sequences[1-4] such as T1w [1] and T2w[4], and specialised bone-imaging sequence such as UTE and ZTE[2], but have yet to use black-bone sequences. A short echo-time also works to reduce susceptibility-induced geometric distortion, making the sequence well suited for radiotherapy (RT) planning. This work aims to evaluate whether the intrinsic CT-like appearance of inverted black-bone MRI can be exploited for effective pelvic DL-based sCT generation, demonstrating sequence feasibility.
Methods
An optimised PD-weighted 3D-GRE StarVIBE (TE=4.77, TR=6.4ms, FA=3°) black-bone MRI sequence with deep-resolve, acquired during routine whole-body scanning on a 1.5T Sola scanner, had its contrast inverted and was registered to a paired CT-scan. A 2-stage rigid and affine transformation was applied to register the MRI to the fixed-CT. U-Net and CycleGAN architectures were trained for sCT generation networks using a subject-level train-validation-test split of 7-3-3. RT treatment plans were created using identical structure-sets and beams, and differences in dose to the PTV were reported.
Results
The average MAE for the U-Net and CycleGAN for 3 test-patients was 93.79±14.13 HU and 68.65±3.12 HU, respectively. For a plan made on a U-Net test-patient, the dose to the PTV had a difference of 0.3% when comparing CT/sCT plans. For the CycleGAN, three plans were created. The average dose-difference to the PTV was 0.47±0.14%, with a maximum absolute deviation of 0.63%.
Conclusion
HU errors were comparable to similar literature[1], demonstrating the effectiveness of black-bone MRI in sCT generation. Calculated doses in plans showed clinical equivalence between using CT and black-bone based sCT.