4D-MRI for thoracic radiotherapy treatment planning and guidance on an MR-linac

Abstract

Lung cancer currently has a poor prognosis in England and Wales with only 10% of adults surviving five years or more after diagnosis. The poor survival rate demonstrates the need to improve contemporary treatment approaches such as external beam radiotherapy, which is part of the treatment course for approximately 30% of lung cancer patients. In thoracic radiotherapy workflows, the tumour site is treated with targeted radiation over several treatment sessions (fractions). Conventionally, each fraction is planned based on a single 4D-CT scan acquired prior to the first fraction. This planning approach does not account for changes in patient anatomy or respiratory pattern occurring throughout the treatment course, which can cause unnecessary irradiation of healthy tissue and under-dosage of the tumour site. Inter-fractional changes in anatomy and respiratory pattern could be corrected for by adapting the treatment plan using 4D-MRI acquired at the beginning of each fraction. It is feasible to perform MR imaging, plan adaptation and treatment delivery in the same fraction using an MR-linac system. In this studentship, methods to obtain 4D-MRI for thoracic radiotherapy treatment planning and guidance on an MR-linac were explored. In particular, methods to calculate T2-weighted MRI (4D-T2w MRI) and synthetic 4D-CT (4D-sCT) were investigated. Compared to 4D-CT and T1-weighted 4D-MRI, the high soft-tissue contrast exhibited by 4D-T2w MRI could facilitate delineation of tumour sites affected by respiratory motion, especially when adjacent to areas of healthy tissue. Two methods to obtain 4D-T2w MRI were developed and verified. In the motion vector field projection (MVFP) method, 4D-T2w MRI were calculated by applying the motion information from 4D-T1w to 3D-T2w MRI. In the super-resolution approach, continuously acquired axial and sagittal 2D-T2w images were combined into one high-resolution 4D-T2w image using binning, image registration and super-resolution reconstruction. Unlike existing techniques based on slice-selection, the MVFP and super-resolution methods resulted in geometrically accurate 4D-T2w MRI with high spatio-temporal resolution. 4D-sCT can provide motion-compressed electron density information required for treatment planning. Three consecutive techniques (Dixon-based, Dixon-Spine and Dixon-Spine-Lung) to generate 4D-sCT were developed and validated. In the Dixon-Spine-Lung method, 4D-sCT were obtained using blk-density assignment (Dixon), best-atlas segmentation (spine), polynomial fitting (lung) and motion-modelling (4D component). Overall, good dosimetric agreement was found between Dixon-Spine-Lung 4D-sCT and 4D-CT. Prior to this studentship, no MRI-derived 4D-sCT implementation was published. Deep learning was utilised to accelerate motion-compensated 4D-T1w MRI reconstruction from 9-12 hours to 28 seconds. The MVFP method was accelerated using deep learning-based 4D-T1w MRI and then applied to obtain 4D-T2w MRI from MR-linac data. Compared to super-resolution reconstructed and phase binned 4D-T2w MRI, MVFP-generated 4D-T2w MRI was found best suited for on-line application in an MR-linac workflow. In particular, it exhibited good spatio-temporal resolution and required less than 6 minutes to reconstruct.