Feasibility of Magnetic Resonance only Radiotherapy for Patients with Gynaecological Carcinoma

Abstract

In conventional radiotherapy planning pathways, the CT scan provides both anatomical and electron density information required for accurate dose calculation. The poor soft tissue contrast associated with CT images is traditionally overcome through fusion of MRI images with the radiotherapy planning CT. The registration of the MR images with the planning CT brings its own uncertainties. To overcome this, there has been increased interest in generating synthetic CT (sCT) datasets directly from the MRI scans. The purpose of this work was to generate a female pelvic model for sCT generation using a deep learning solution based on generative adversarial networks (cycleGAN). The model is trained on 30 MR-CT paired datasets incorporating standard T2 spin echo scans. A further 10 patients were used to test the model using DVH parameter comparison, HU mean average error and gamma analysis. Comparisons were made between a CT deformed to the geometry of the MR dataset (dCT) and the MR derived sCT. Patient contours and clinical targets were generated using artificial intelligence based modelling (MVision) with Planning Target Volumes generated using standard clinical margins. CT numbers for bony structures in the sCT were lower compared with the planning CT whilst the CT numbers for adipose and soft tissue were comparable. Dose differences between the dCT and sCT were 0.2% (D98%). Gamma analysis showed a mean of 90.4% at 1%/1mm. ‘End to End’ (E2E) testing of processes, specifically dosimetric validation, is a prerequisite for efficient and safe implementation of any novel and/or complex radiotherapy technique. For this purpose, several aspects of the patient workflow have been tested including delivery of the treatment plans using a linear accelerator using techniques that interrogate the distribution in 1 Dimension (chambers), 2D (film and Delta4) and 3D (gel dosimetry). E2E testing in a 3D printed anthropomorphic phantom (RTsafe) using a PTW 0.125cc semiflex chamber, EBT3 and EBT4 film and RT100 gel was performed. The mean dose difference between the planned dose and the semiflex dose was 0.76 % For the EBT4 film, this difference was -0.8 % in the coronal plane and 0.5 % in the sagittal plane. Gamma analysis using the delta4 ranged from 88.2% to 97.8% for 3%/2mm. One dimensional gamma analysis using the gel ranged from 59.2% to 100% at 5%/2mm. The study also considered the accuracy of using MR for the treatment verification element of image guided radiotherapy using the rigid registration method in the treatment planning system. The mean differences between the MR-CBCT and sCT-CBCT matching (as compared to the standard CT-CBCT matching) were the same regardless of whether the MR or sCT was used though the standard deviation was higher for the sCT-CBCT matching. Assuming MR can be implemented, without detriment to data integrity, as the primary reference image in the oncology management system, it should be used for IGRT. This study has shown that acceptable sCT can be produced using standard T2 sequences using deep learning for female pelvis. The accuracy of the doses calculated has been verified using both treatment planning system comparisons and external measurements of dose delivery. Either the sCT or the MR could be used to provide a reference image for IGRT. Finally a critical review of the work has been undertaken and recommendation for future work towards implementation have been provided.

Date of Award	2 Jul 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Robert Chuter (Main Supervisor) & Michael Taylor (Co Supervisor)