2793 - Deep Learning-Based Synthetic CT Generation from MRI for Head and Neck Cancer Radiation Therapy Planning
Presenter(s)
D. Liefman1,2, P. Ramachandran1,3, C. Jones1, C. Gould1, A. Devlin1, L. J. McDowell1,3, and H. Liu1,3; 1Princess Alexandra Hospital, Brisbane, Australia, 2Bond University, Gold Coast, Australia, 3University of Queensland, Brisbane, Australia
Purpose/Objective(s): We present a novel deep learning (DL) model trained to generate synthetic computed tomography (sCT) from magnetic resonance imaging (MRI), with accurate Hounsfield Unit (HU) values acceptable for radiation therapy (RT) dosimetric planning. An MR-only workflow eliminates MRI-CT registration error and integral ionizing radiation dose, improving the patient experience by reducing total time spent immobilized in a thermoplastic mask.
Materials/Methods: 60 adult patients treated with curative intent RT for oropharyngeal cancer between January 2023 & January 2024 were identified. 50 imaging datasets were used for DL model training & 10 for validation. Reflecting real-world populations, 25 training and 2 validation cases had dental implants. Datasets with extensive motion artefact or reconstructive prostheses were excluded. Patients underwent a non-contrast planning CT (pCT) (Toshiba Aquilion Large Bore, 2 mm slices) and a head and neck (H&N) MRI (a technology company's full-body MRI scanner, axial 3D T1W Dixon post-gadolinium, 0.9 mm slices) with field of view (FOV) from hard palate to T4 vertebra. Manual then automated rigid image registration was performed in oncology imaging informatics system (a technology company), then single-pass deformable registration using mutual information metric.
The DL model was trained in Python (v3.12.7) using paired MRI-CT datasets (50 patients, 10000 images), applying intensity normalization. sCT images were generated using a UNet-based diffusion model (50 epochs), validated with 10 new datasets (2000 images). HU value agreement between sCT and ground truth pCT was evaluated using standard quantitative comparative metrics Dice Similarity Co-efficient (DSC), Normalized Cross Correlation Coefficient (NCC), peak Signal-to-Noise Ratio (pSNR) & a novel frequency domain space metric (FDS), with each metric reported 0 to 1 (0 signifying no correlation & 1 signifying perfect HU correlation between sCT & pCT).
Results: The trained DL model generated sCT from MRI input with excellent HU agreement to pCT. The mean values in the validation cohort were as follows: DSC 0.89 ± 0.06, NCC 0.98 ± 0.02, pSNR 0.99 ± 0.01 and FDS 0.99 ± 0.01, with values 0.85-0.99 signifying accuracy appropriate for clinical imaging and dosimetric uses. Compared with other publications, a higher number of paired MRI-CT images and epochs in DL training were used.
Conclusion: This novel DL model developed for generating sCT from H&N MRI yielded HU values highly correlative with pCT on validation. Evaluation of sCT generated from T2W MRI alone will be undertaken next, allowing omission of intravenous contrast administration. MR-only planning workflows are in use internationally for RT treatment of intracranial and genitourinary cancers. Soft tissue discrimination on MRI is particularly advantageous in H&N cancer RT planning. Creating an MR-only workflow for H&N patients could facilitate adaptive planning applications, further optimizing tumor control and reducing toxicity.