Development and Validation of a 2D/3D Registration Software Framework for Patient Positioning in Radiotherapy

Abstract

Philipp Steininger

Although ionizing irradiation per se is not perceptible by human eyes, targeted beam delivery in radiation therapy is similarly efficient as well as critical as surgical interventions by scalpel. Both the matters of fact and provocation included in this sentence have driven scientific and industrial efforts to develop and validate new technologies in the field of image guided radiotherapy (IGRT). In this context, one of the doubtlessly most pertinent tools of medical software algorithms is image registration, which enables spatial tracking of the patient’s anatomy before and during treatment in order to guarantee precise irradiation of cancerous cells, and to simultaneously spare healthy tissue.

This thesis describes the design, implementation and validation of an image registration algorithm, which is capable of rigidly aligning a pre-interventional 3D reference computed tomography (CT) with one or more intra-interventional 2D x-ray projections of the patient. This approach, referred to as 2D/3D image registration, has not only been designed as academic proof of concept, but also to consider real-world conditions in clinical routine. A special focus was set on concepts which counteract alignment uncertainties introduced by potentially moving body parts and occultation emerging from arbitrary objects such as immobilization devices. Moreover, clinical applicability of the developed software has been emphasized by providing an automated implementation, which is seamlessly integrated into clinical workflow with reasonable computational performance.

The resultant software framework, namely the universal n-way open 2D/3D registration (UNO-2-3-REG), was developed according to standardized software development process guidelines, which opened the doors for translation into clinics as part of the medical product “open-radART”. As such, UNO-2-3-REG had to undergo a broad clinical validation prior to release. Beside of elaborate design, implementation and verification of the 2D/3D registration algorithm and connected concepts, one of the major achievements was to validate accuracy of the technique in comparison to interfractional 3D/3D IGRT based on gold standard cone-beam CT (CBCT) workflows in pelvic treatment. Over more than 8000 registration experiments UNO 2 3 REG along with a novel auto-masking strategy to exclude the potentially moving femora were applied to register clinical reference CTs with orthogonal x-ray image pairs of CBCT raw data. Resultant rigid alignment parameters were compared with the results from 3D/3D CBCT-based registration of the according data sets. The algorithm featured translational errors <0.18 ° and rotational errors <0.73 mm in root mean square. Remaining residual errors have mainly been attributed to patient motion on the one hand, and inadequate geometric calibration of the onboard imaging system on the other hand.

From these results we conclude that the presented 2D/3D registration approach based on sequentially acquired orthogonal x-rays of the pelvis is a viable alternative to cone beam CTs without any clinically relevant alteration of geometrical precision. In all cases where rigid alignment on bony anatomy is clinically sufficient for daily IGRT, no time-consuming volumetric intra-interventional data set acquisition is required.
The developed methodology and gained conclusions pose a solid basis for further research and translational work, especially in the field of multi-component 2D/3D image registration and automated region of interest determination.