Author: Gregory Bolard, Rabten Datsang, Sarah Ghandour, Timo Kiljunen, Pauliina Paavilainen, Sami Suilamo, Katlin Tiigi 👨🔬
Affiliation: Turku University Hospital, Virginia Commonwealth University, MVision AI, North Estonia Medical Centre, Docrates Cancer Center, Hopital Riviera-Chablais 🌍
Purpose: To verify the performance of a vendor-neutral deep learning model for synthetic CT generation from T2-weighted and balanced steady-state MR sequences to support both MR-only simulation and MR-guided radiotherapy workflows for pelvic soft tissue cancers.
Methods: A CNN-based model (Image+ Pelvis), trained on 485 paired MR-CT 3D datasets with domain-specific augmentation, was validated on 45 pelvis datasets from multiple MR platforms including conventional simulators (GE Signa 1.5T, n=5; Philips Ingenia 1.5T, n=22; Siemens MAGNETOM Vida 3T, n=5) and MR-Linacs (Elekta Unity, n=7; ViewRay MRIdian, n=6). Input sequences included T2-weighted spin echo (2D/3D) and balanced steady-state sequences with 40-50 cm FOV and sub-3mm slice thickness. Clinical VMAT/IMRT plans (6X, 6XFFF, 7XFFF) for various pelvic cancers (prostate, rectum, cervix, bladder) were recalculated using SciMoCa™ Monte Carlo (2mm grid, 0.5% uncertainty). Performance metrics included 3D gamma analysis (2%/2mm, 3%/3mm), dose differences in targets and organs at risk (OARs).
Results: Quantitative analysis showed excellent agreement between synthetic and planning CT-based calculations (linear regression: 1.003, R²=0.999 for mean doses). Target volumes (n=75) showed mean relative dose differences of 0.33% ± 0.40%. For OARs, differences ranged from -0.08% ± 0.73% (rectum) to 0.45% ± 0.35% (pelvic bone). Gamma analysis demonstrated high pass rates at 2%/2mm (99.2% ± 1.1% low dose, 97.5% ± 3.1% high dose regions) and 3%/3mm (>99.8%). One case showed lower performance due to strong motion artifacts in input MRI (2%/2mm: 78.3%), and another showed -1.0 Gy rectum dose difference due to air filling variations (air cavities not predicted by the model).
Conclusion: This study verifies clinically acceptable dosimetric accuracy across MR platforms for photon treatment of soft tissue cancers in the pelvis area. The model's robust performance supports its integration into MR-only and adaptive MR-guided radiotherapy workflows. MR image quality remains an important aspect for synthetic imaging quality.