The aim of adaptive radiotherapy is to reduce dose deviations from the initial irradiation plan that may result from geometric uncertainties. Prior to the start of radiotherapy, the so-called planning images of the patient are first acquired using computed tomography (CT), magnetic resonance (MR) and/or positron emission tomography (PET), then segmentations of structures of interest, i.e. targets (tumors) and organs-at-risk, and treatment plans are created based on these images. Based on the plan, the radiotherapy is carried out across several daily fractions. Because geometric uncertainties are introduced during treatment in the form of anatomical changes and changes in the patient's positioning, it is necessary to acquire new images prior to each fraction and adapt the treatment plan based on these images. Pre-fraction images are often acquired using cone-beam computed tomography (CBCT) because such imaging is performed in-room, while the patient is immobilized. The low-dose CBCT imaging exhibits a lower image quality, which complicates the procedures of reliable automatic segmentation of the structures and adjustments of the treatment plan. An alternative approach is to use high-quality planning images and register them onto the CBCT images, and then map the segmentations into the CBCT image space. Because structures in the body often deform and move independently of each other, a type of non-rigid registration called deformable image registration (DIR) is required. This thesis gives a general description of the workings of DIR and lists some of its uses in adaptive radiotherapy. The aim was to find, study and quantitatively and comparatively evaluate different algorithms for deformable image registration. To this end we employ datasets of sixty pairs of CT and CBCT images of prostate cancer patients, acquired during an actual radiotherapy, and two image pairs from the public TG-132 database in order to assess the success of DIR algorithms. Quantitative evaluation was based on the following metrics: (i) surface and volumetric Dice coefficients between segmentations of the patient skin, prostate and left femur, (ii) distance between ten corresponding landmarks of each image pair using target registration errors (TRE), (iii) four similarity measures computed before and after application of DIR to each image pair, and (iv) the count of voxels with negative determinant of the Jacobian matrix (i.e. voxels with physically implausible deformation). The results are presented in three parts. In the first part we validated the pTVreg, Elastix, Plastimatch and Deeds registration algorithms with experimentally determined parameters. In the second part we introduced a robust way of finding optimal registration parameters with the design of experiments method using Taguchi matrices (DOE), and then compared the results to previous ones. In the third part, we further optimized the parameters to speed-up DIR with the Deeds algorithm, which to this stage showed the best registration results. In the end, we showed that the Deeds algorithm proved the best in terms of accuracy and robustness, as it achieved the best results during evaluations with the Dice coefficients and TRE. The analysis of the determinants of the Jacobian matrices showed that the Deeds algorithm exhibited little or no folding. On the virtual phantom image pair from the TG-132 dataset all tested algorithms achieved the recommended values of Dice coefficients, and on the TG-132 thorax CT image pair from the DIR-Lab dataset, the recommended values of TRE were only achieved by the Deeds algorithm, with an average value and standard deviation of 1.69 ± 2.37 mm. With the use of DOE methodology, we also significantly reduced the registration time of one image pair with the Deeds algorithm from 433 seconds to 249 seconds without any noticeable loss in registration accuracy and robustness.
|
