Multicentre, deep learning, synthetic-CT generation for ano-rectal MR-only radiotherapy treatment planning

Background and purpose: Comprehensive dosimetric analysis is required prior to the clinical implementation of pelvic MR-only sites, other than prostate, due to the limited number of site specific synthetic-CT (sCT) dosimetric assessments in the literature. This study aims to provide a comprehensive assessment of a deep learning-based, conditional generative adversarial network (cGAN) model for a large ano-rectal cancer cohort. The following challenges were investigated; T2-SPACE MR sequences, patient data from multiple centres and the impact of sex and cancer site on sCT quality.

Method: RT treatment position CT and T2-SPACE MR scans, from two centres, were collected for 90 ano-rectal patients. A cGAN model trained using a focal loss function, was trained and tested on 46 and 44 CT-MR ano-rectal datasets, paired using deformable registration, respectively. VMAT plans were created on CT and recalculated on sCT. Dose differences and gamma indices assessed sCT dosimetric accuracy. A linear mixed effect (LME) model assessed the impact of centre, sex and cancer site.

Results: A mean PTV D95% dose difference of 0.1% (range: -0.5% to 0.7%) was found between CT and sCT. All gamma index (1%/1 mm threshold) measurements were >99.0%. The LME model found the impact of modality, cancer site, sex and centre was clinically insignificant (effect ranges: -0.4% and 0.3%). The mean dose difference for all OAR constraints was 0.1%.

Conclusion: Focal loss cGAN models using T2-SPACE MR sequences from multiple centres can produce generalisable, dosimetrically accurate sCTs for ano-rectal cancers.