Synthetic CT imaging for pelvic and abdominal MR-only radiotherapy: Clinical validation on a 1.5T MR-Linac.

BACKGROUND

Magnetic Resonance (MR)- only workflows are fundamentally limited by the absence of electron density information required for accurate photon dose calculations, necessitating supplementary CT imaging. Synthetic CT (sCT) generation algorithms offer a promising solution to eliminate this dependency. While several commercially available sCT solutions exist, their clinical scope remains highly restricted to specific MRI sequences and anatomical regions.

PURPOSE: To evaluate two deep learning-based sCT generation models for pelvic and abdominal anatomies using both T1 and T2-weighted MR sequences. This study aims to validate their potential for enabling streamlined MR-only radiotherapy workflows eliminating the need for supplementary CT imaging while reducing patient burden and workflow complexity.

METHODS

31 patients with multiple MRI sequences underwent sCT generation using two deep learning models (Image+, MVision AI). Geometric accuracy and Hounsfield Unit (HU) fidelity were assessed through quantitative comparison with deformed CTs (dCT) registered to MRI geometry. Dosimetric validation was performed by comparing dose distributions using γ-index pass rates between sCT-based plans and two reference standards: deformed CT (dCT) and clinical bulk-density CT (bCT).

RESULTS

Mean Dice similarity coefficients demonstrated good geometric agreement: 0.73 ± 0.18 (air cavities), 0.86 ± 0.06 (adipose tissue), 0.85 ± 0.04 (soft tissue), and 0.75 ± 0.07 (bone). Dosimetric evaluation revealed averaged relative dose differences of 0.31% ± 0.98% (bCT reference) and 0.16% ± 1.13% (dCT reference) across anatomical structures. Clinical dose agreement within 2% was achieved in 98.8% of pelvic cases and 93.3% of abdominal cases.

CONCLUSION

MRI-based sCTs from typical T1 or T2-weighted MR linac sequences demonstrated strong geometric and dosimetric agreement with dCT, particularly in the pelvic region, with performance appearing independent of MR sequence type. These results support the clinical viability of sCTs as an alternative to conventional CT, establishing the feasibility of comprehensive MR-only RT workflows from simulation through treatment delivery.