Reshma Munbodh - Univeristy of Pennsylvania - Sub-millimetre 2D-3D registration for image-guided radiation therapy

Cancer patients usually undergo external beam radiation therapy whereby high doses of ionising radiation are delivered to the tumour either in a fractionated scheme or as a single large fraction. Techniques such as 3D conformal radiation therapy and intensity modulated radiation therapy allow accurate specification of the target volume to be irradiated in 3D and the design of highly conformal treatment plans that incrementally spare normal tissues thereby enabling an escalation in radiation dose, improved local tumour control and reduced complications. These methods, however, are highly dependent on accurate treatment delivery. Geometric uncertainties in radiation delivery may result in under-dosage of the tumor leading to local tumour recurrence or unacceptable morbidity from over-dosage of neighbouring healthy tissues. It is essential, therefore, to ensure that the patient is set up accurately during every treatment fraction. One approach towards this is to employ an image-guided verification procedure that uses 2D projection radiographs, acquired during treatment at kilovoltage (kV) or megavoltage (MV) photon energies, and a 3D kV planning CT or 3D kV cone beam CT (CBCT) image, in conjunction with conventional immobilisation techniques to correct patient setup errors prior to radiation delivery. We will show that sub-millimetre estimation of patient setup errors is achievable with multiple methods within an automated 2D-3D registration framework based on an understanding of the physics of transmission images. The 2D-3D registration methods we present address two sources of noise in the images: low-frequency noise due mainly to scattered radiation and Poisson noise. The first registration method we present uses intensity-derived features, representing bony ridges, within an intensity-based registration framework and is robust to the low signal-to-noise ratio of 2D radiographs. The other methods we present are based on a statistical model of the intensity values within the two imaging modalities being registered. The model assumes that intensity values in projection radiographs are independently but not identically distributed due to the non-stationary nature of photon counting noise. Using maximum likelihood estimation, two similarity measures, maximum likelihood with a Poisson distribution and maximum likelihood with a Gaussian distribution are derived. Further, we investigate the merit of the model-based registration approach for data obtained with current imaging equipment and doses by comparing the performance of the above methods to that of intensity-based registration using the Pearson correlation coefficient on accurately collected data of anthropomorphic phantoms of the pelvis and the head and on patient data. The value of the solutions presented lie in their automated nature, accuracy and dependability as well as their crucial role towards the development of closed-loop treatment procedures that allow for tighter treatment margins, reduced toxicity and increased patient safety during radiation therapy.