Development of an Image Correction Pipeline for Upright Radiotherapy Delivered via a Conventional Linac

Radiation therapy is a standard treatment for cancer patients, typically delivered to a recumbent patient. However, this position can cause discomfort and breathing difficulties for some patients. Treatment of patients in an upright position has the potential to alleviate these issues, which could improve patient stability, enable treatment where it might not otherwise be possible, and improve treatment geometry for some patients. This thesis aims to develop a pipeline for delivering upright radiotherapy using a standard Linear Accelerator (linac) for imaging and treatment, without requiring specialised equipment. Acquiring a high-quality upright 3D image is critical in upright radiation therapy. While vertical CT scanners exist, this thesis uses the linac onboard imaging (OBI) to acquire upright cone-beam computed tomography (CBCT) scans, reducing the need for specialised equipment. However, due to lower image quality, CBCT scans are not commonly used for treatment planning. This thesis focuses on developing in-house image corrections (both optimising published methods and applying novel solutions) to make the CBCT images suitable for treatment planning by increasing the CT number accuracy. The image corrections are developed using standard supine CBCT images with an in-house reconstruction pipeline constructed. The pipeline is adapted so that acquired upright CBCT images can be corrected and reconstructed. This allows upright CBCT treatment plans to be generated and delivered via a conventional linac. To improve the image quality of CBCT images, in-house correction techniques (for scattered radiation, spatial variation in CT number induced by the bowtie filter spectral changes, and ring artefacts) have been developed. The goal was to improve CT number accuracy so that the images could be used for accurate dose calculation in the treatment planning process (since the calculated uses voxel electron density, calculated from the CT number using a CT-ED curve). (since this calculation uses the electron density of a voxel, which is proportional to the CT number). While other artefacts were corrected, the corrections' primary focus is on scattered radiation, which significantly alters CT numbers. Three methods have been tested. They have been divided based on their need for a Fan Beam Computed Tomography (FBCT) image, which is typically not compatible with upright imaging. The first is a Monte Carlo (MC) method that requires an FBCT image as an accurate virtual model of the patient in the scatter simulation process and subtracts scatter from each projection using simulated scatter maps. The second method is an iterative MC method that uses the CBCT image to simulate scatter maps iteratively and does not require an FBCT image. The final method, which does not require an FBCT image, generates scatter-free projections using a specially trained neural network (NN). It performs similarly to the FBCT-based MC method for anthropomorphic data (as it resembles its training data) and is faster than the other two methods. All three correction techniques are applied to the CBCT projections. Additionally, two other corrections (for the effects of spatial variation in CT number induced by the bowtie filter and bowtie filter spectral changes and ring artefacts) are applied to the reconstructed image. The former method segments the image by material. It modifies the linear attenuation by considering the pixel position corresponding to the thickness of the bowtie filter portion, allowing for a material and position-dependent correction. The second correction technique corrects ring artefacts by transforming the image into polar coordinates (making the ring artefacts line artefact), allowing for a correction to be calculated and applied per row (leaving rows lacking the artefact unaffected). The linac OBI and an independent turntable are used to acquire upright CBCT projections. The upright CBCT images had worse image quality than supine CBCT images (MAE from the FBCT image of 102 HU compared to 61 HU), both reconstructed via the in-house pipeline. This is due to the lack of geometry calibration and a bowtie filter, as well as a noisier flood field. Due to not requiring a FBCT image, both the iterative and NN methods were applicable for upright image application. Among the tested methods, the NN method demonstrated the most significant improvement in image quality (MAE of 78 HU compared to 90 HU and 88 HU for the MC-based methods), indicating that it is the recommended approach for upright imaging. However, further work is required to improve the quality of the uncorrected upright images. Treatment plans generated from corrected CBCT images have smaller dose differences from the dose planned on the FBCT image than the original CBCT images. Consistent with reproducing the CT numbers most accurately, the FBCT-based and NN corrections correspond to the highest dose difference passing rates. While the NN correction does not perform as well when beam hardening causes information loss, the results support its use when the FBCT-based method is not applicable. Treatment plans generated on corrected upright CBCT images differ from those generated from FBCT images due to reduced upright image quality and CT number accuracy. Doses measured from both supine and upright deliveries compared similarly to their respective planned doses (2%2mm gamma passing rate of 88% and 80%, respectively). Currently, kV image guidance cannot be used in conjunction with the upright delivery due to the position of the treatment couch. Before upright patient treatment, image guidance must be appropriately adapted to the upright orientation. The results suggest that using corrected CBCT images to plan upright treatment is feasible, but more research is needed. This thesis focuses on developing an image correction pipeline that produces upright CBCT images with a standard linac and the only specialised equipment, a turntable, to position and rotate the patient. Although there are currently limitations in acquiring upright CBCT images, the current pipeline can correct and reconstruct the projections. However, additional components, such as geometrical calibration, are required to improve the process further. Despite the need for improvement, upright treatment plans calculated on corrected upright CBCT images can be delivered without significant dose differences, indicating the process's potential for success. This thesis lays the groundwork for future clinical trials. The anticipated long-term outcome of this research is to provide upright radiotherapy as a standard clinical treatment option.