dc.description.abstract
Cancer is the second most common cause of death in the world, responsible for approximately 10 million deaths in 2019. Survival rates have improved significantly in the last 50 years, and the increase in life expectancy after a radiation therapy resulted in an increase in the incidence of recurrences and new cancers in people who have already undergone radiotherapy in the past. Hence, there is a growing usage of re-irradiation in recent practice. For patient safety, it is necessary to use registration in order to localize previously irradiated tissue from a former treatment and consider it as background of the new treatment. However, there is no common standard or protocol for re-irradiation and a lack of clinical consensus, as well as common ground for the studies done on this subject. Moreover, the evaluation of dose mapping is complicated as there is no ground truth to compare results to. This thesis uses data from the ReCare trial to explore possible data analysis strategies and aims to calculate the best estimate of dose deformation by combining different registrations, and derive dosimetric uncertainties from it. It was also an additional goal of the project to determine whether it would be possible to use an AI-powered segmentation tool as QA for the manual registrations coming from different institutions in the ReCare cohort."Whole body" segmentations were performed with TotalSegmentator on 3D Slicer and compared with the manual segmentations on RayStation with HD and DSC metrics. To increase the number of organs to compare, the AI-colon and AI-spinal cord were modified to a pseudo-AI rectum and a pseudo-AI cauda equina, matching the ReCare segmentations. On RayStation, a rigid image registration (RIR) was performed, and, using ANACONDA, an intensity-based deformable image registration (DIR) (noCont- ROI), as well as two hybrid DIRs (ContROI & Wall). On 3D Slicer, an intensity-based DIR was performed using Elastix (SDIR). The calculation of dosimetric uncertainties in dose deformation was based on a best estimate made of the five registrations, assuming that a geometry-based registration could deform dose because of the spatial correlation between anatomy and dose distribution. The highest 1% and average standard deviation of dose were recorded for each organ-at-risk (OAR) and the PTVs of the second treatments.A very good similarity (lowest Dice similarity coefficient (DSC): Rectum 0.79) was found between AI and manual segmentation. However, a low structure correspondence was found: only three organs are segmented by both methods, whereas the list of OARs in the ReCare trial counts ten of them. Moreover, the theoretical time gain could not be made profitable because of the capacities of the computer used. The general DSC means retrieved were 0.81 for ContROI, 0.74 for noContROI, and 0.70 for SDIR. The general HD means measured were 1.57 for ContROI, 1.87 for noContROI and 1.69 for SDIR. DIR performed particularly better for filling organs such as the bladder, the bowel and the rectum, which all exhibited significant differences between algorithms. Significant DSC differences were also found between the two intensity-based algorithms, ANACONDA gave better DSC values, while Elastix performed better for HD measurements. Dosimetric uncertainties were found to be generated by a complex combination of geometric uncertainties, a steep gradient dose distribution, and a certain dose magnitude. Geometric variations were found to be due to image information, to the use of controlling structures, and to differences in the handling of body contours. For intensity-caused variations, the anal canal region was identified as being prone to dosimetric uncertainties. The presence or absence of the RIR in the calculation of dosimetric uncertainties was not found to be significant in the majority of OARs or in the PTVs. The PTVs’ uD,av range from 0.30% to 13.90%, which highlights how case-specific the impact of dosimetric uncertainties is on the patient’s safety.The lack of structure correspondence between manual and AI-segmentation limits the use of TotalSegmentator as a QA tool for the ReCare manual segmentations. The registration analysis resulted in the observation that hybrid DIR gives better results than intensity-based DIRs, as it offers more subjectivity to match the specificity of each case. The differences between the two intensity-base DIRs were found to be due to their use of different optimization metrics. The significant dosimetric uncertainties in the body contour were attributed to the inability of the RIR to correct for them. Quantifying and localizing dosimetric uncertainties, such as those extracted in the new treatment PTVs, demonstrates how incorporating these uncertainties into treatment planning can potentially increase patient safety. Further research using additional patients is necessary to strengthen the results, confirm the trends identified, and deepen the analysis of factors of dosimetric uncertainties. Additionally, examining other anatomical regions where TotalSegmentator might be more suitable and DIR algorithms could yield different results would be an interesting direction for further investigations.
en
