[en] Ths paper examines several applications of deformable registration algorithms in the field of image-guided radiotherapy. The first part focuses on the description of input and output of deformable registration algorithms, with a brief review of conventional and most current methods. The typical applications of deformable registration are then reviewed on the basis of four practical examples. The first two sets of examples deal with the fusion of images obtained from the same patient (inter-fraction registration), with time intervals of several days between each image. The other two examples deal with the fusion of images obtained in immediate sequence (intra-fraction registration); in this case, the focus is the displacement during image acquisition or patient treatment (mainly due to respiratory movement), with time intervals in the order of magnitude of tenths of seconds. Finally, the registration of images of different patients (inter-patient registration) is also discussed. In conclusion, deformable registration has become a fundamental tool for image analysis in radiotherapy. Although extensive validation of the numerous existing methods is required before extending its clinical use, deformable registration is expected to become a standard methodology in the treatment planning systems in the near future. (orig.)

[en] To be efficient, the treatment of the lung cancers with radiation therapy must take into account the respiratory motion. The knowledge of this motion requires the acquisition of 4-dimensional computed tomography (4D CT) images. The free-breathing thorax 4D CT images currently acquired use gated or respiratory-correlated methods. These methods involve the collection of a respiratory signal during the acquisition of data in order to sort them into different groups. The quality of the 4D CT image thus depends on an accurate description by the signal of the position of the thorax in the respiratory cycle. The signal is generally acquired by independent measurements of densitometric data (spirometer, thermometer,...). We propose to extract it directly from the sequence of 2-dimensional cone-beam (2D CB) projections acquired around the free-breathing thorax. Our method derives the motion between two consecutive 2D CB projections by using a block matching algorithm. Blocks are positioned around points of interest constituting a regular sampling of the 2D CB projections. A unidimensional signal is derived from the trajectory of each block in the sequence after projection. The aggregation of a subset of selected signals makes it possible to derive the respiratory signal during the acquisition time. Our method is validated quantitatively on simulated data and qualitatively on real data. On simulated data, we obtain a respiratory signal with 97.5 % linear correlation with the reference. On real data, the extracted signal allows us to reconstruct the 4D CT image and compare it with the blurred 3D CT image obtained without taking into account the respiratory motion. (authors)

[en] A method is proposed and evaluated to model large and inconvenient phase space files used in Monte Carlo simulations by a compact generative adversarial network (GAN). The GAN is trained based on a phase space dataset to create a neural network, called Generator (G), allowing G to mimic the multidimensional data distribution of the phase space. At the end of the training process, G is stored with about 0.5 million weights, around 10 MB, instead of a few GB of the initial file. Particles are then generated with G to replace the phase space dataset.

This concept is applied to beam models from linear accelerators (linacs) and from brachytherapy seed models. Simulations using particles from the reference phase space on one hand and those generated by the GAN on the other hand were compared. 3D distributions of deposited energy obtained from source distributions generated by the GAN were close to the reference ones, with less than 1% of voxel-by-voxel relative difference. Sharp parts such as the brachytherapy emission lines in the energy spectra were not perfectly modeled by the GAN. Detailed statistical properties and limitations of the GAN-generated particles still require further investigation, but the proposed exploratory approach is already promising and paves the way for a wide range of applications. (paper)

[en] A method is proposed to model by a generative adversarial network the distribution of particles exiting a patient during Monte Carlo simulation of emission tomography imaging devices. The resulting compact neural network is then able to generate particles exiting the patient, going towards the detectors, avoiding costly particle tracking within the patient. As a proof of concept, the method is evaluated for single photon emission computed tomography (SPECT) imaging and combined with another neural network modeling the detector response function (ARF-nn). A complete rotating SPECT acquisition can be simulated with reduced computation time compared to conventional Monte Carlo simulation. It also allows the user to perform simulations with several imaging systems or parameters, which is useful for imaging system design. (paper)

[en] A method to speed up simulations of single photon emission computed tomography (SPECT) imaging is proposed. It uses an artificial neural network (ANN) to learn the angular response function (ARF) of a collimator–detector system. The ANN is trained once from a complete simulation including the complete detector head with collimator, crystal, and digitization process. In the simulation, particle tracking inside the SPECT head is replaced by a plane. Photons are stopped at the plane and the energy and direction are used as input to the ANN, which provides detection probabilities in each energy window. Compared to histogram-based ARF, the proposed method is less dependent on the statistics of the training data, provides similar simulation efficiency, and requires less training data. The implementation is available within the GATE platform. (paper)

[en] Monte-Carlo simulations of SPECT images are notoriously slow to converge due to the large ratio between the number of photons emitted and detected in the collimator. This work proposes a method to accelerate the simulations based on fixed forced detection (FFD) combined with an analytical response of the detector. FFD is based on a Monte-Carlo simulation but forces the detection of a photon in each detector pixel weighted by the probability of emission (or scattering) and transmission to this pixel. The method was evaluated with numerical phantoms and on patient images. We obtained differences with analog Monte Carlo lower than the statistical uncertainty. The overall computing time gain can reach up to five orders of magnitude. Source code and examples are available in the Gate V8.0 release. (paper)

[en] The GATE package is used to perform Monte Carlo hadron-therapy simulations of a cancer treatment combined with the complete description of an associated positron emission tomography (PET) imaging device for dose monitoring. This study aimed to demonstrate that the GATE platform has the capability to perform realistic simulations in the field of hadron-therapy, combining both dose and imaging system. We defined the simulation configuration as a carbon ion pencil beam scanning of a thorax CT phantom together with a complete PET imaging system. Two main positron emitters resulting from nuclear reactions are considered in this study: and we studied the produced data to analyse the interest in using a full PET system simulation instead of the usual Gaussian smoothing applied on the positron emitters map. We found differences in the distal position of the signal falloff of 20% between full PET system simulation and the Gaussian model. We also studied the influence of the isotope in PET images and found the contribution to falloff of this isotope to be negligible (4%), which suggests the inclusion of isotope in the simulation is not necessary. Finally, we analysed the impact of dose delivery on PET image quality and found a difference of 20% on the PET estimation falloff between doses of 10 Gy and 1 Gy. This study shows that GATE, implemented on a computing system with large number of CPUs ( 1000), has the potential to be used for quantitative evaluation of imaging protocols for radiation monitoring. (authors)

[en] Purpose. Using deformable registration methods from a phase two clinical study of air breath control during radiotherapy in patients suffering from severe respiratory insufficiency and non-small cell lung carcinoma. Patients and methods, Between April 2002 and November 2005, 22 patients with severe respiratory insufficiency were treated with curative intent by conformal therapy combined with active breathing control. Results. After a mean of follow-up of 22 months, the local control rate is 28% and the method is feasible despite the severe respiratory insufficiency. However the overall survival is still poor due to metastatic widespread. For the second part of the study, the clinical protocol was also used for two studies using deformable registration methods. In the first study, a deformable registration method has been developed in order to register several breath-hold 3D CT of the same patient acquired at several days of interval. It allowed quantifying the inter-fraction breath-hold reproducibility by analysing the resulting displacement field. For 6 patients, the breath-hold was effective, while for 2 patients, motion greater than 10 mm were detected. The second study aimed to simulate 4D images from 3D breath-hold images. Developing an ad-hoc methodology based on the interpolation of 3D dense deformation fields performed it. The approach has been validated with expert selected landmarks, with accuracy lower than 3 mm. Conclusion. ABC is feasible, even in case of severe insufficiency respiratory syndrome but metastatic widespread disease is still a major challenge even with an acceptable local control rate without serious side effects: regarding the deformable registration method. Such artificial 4D images could allow decreasing the dose need to acquire a full 4D image, to simulate irregular breathing pattern and to be used for 4D dosimetry planning. (author)

[en] Active scanning delivery systems take full advantage of ion beams to best conform to the tumor and to spare surrounding healthy tissues; however, it is also a challenging technique for quality assurance. In this perspective, we upgraded the GATE/GEANT4 Monte Carlo platform in order to recalculate the treatment planning system (TPS) dose distributions for active scanning systems. A method that allows evaluating the TPS dose distributions with the GATE Monte Carlo platform has been developed and applied to the XiO TPS (Elekta), for the IBA proton pencil beam scanning (PBS) system. First, we evaluated the specificities of each dose engine. A dose-conversion scheme that allows one to convert dose to medium into dose to water was implemented within GATE. Specific test cases in homogeneous and heterogeneous configurations allowed for the estimation of the differences between the beam models implemented in XiO and GATE. Finally, dose distributions of a prostate treatment plan were compared. In homogeneous media, a satisfactory agreement was generally obtained between XiO and GATE. The maximum stopping power difference of 3% occurred in a human tissue of 0.9 g cm⁻³ density and led to a significant range shift. Comparisons in heterogeneous configurations pointed out the limits of the TPS dose calculation accuracy and the superiority of Monte Carlo simulations. The necessity of computing dose to water in our Monte Carlo code for comparisons with TPSs is also presented. Finally, the new capabilities of the platform are applied to a prostate treatment plan and dose differences between both dose engines are analyzed in detail. This work presents a generic method to compare TPS dose distributions with the GATE Monte Carlo platform. It is noteworthy that GATE is also a convenient tool for imaging applications, therefore opening new research possibilities for the PBS modality. (paper)

[en] Sliding motion is a challenge for deformable image registration because it leads to discontinuities in the sought deformation. In this paper, we present a method to handle sliding motion using multiple B-spline transforms. The proposed method decomposes the sought deformation into sliding regions to allow discontinuities at their interfaces, but prevents unrealistic solutions by forcing those interfaces to match. The method was evaluated on 16 lung cancer patients against a single B-spline transform approach and a multi B-spline transforms approach without the sliding constraint at the interface. The target registration error (TRE) was significantly lower with the proposed method (TRE = 1.5 mm) than with the single B-spline approach (TRE = 3.7 mm) and was comparable to the multi B-spline approach without the sliding constraint (TRE = 1.4 mm). The proposed method was also more accurate along region interfaces, with 37% less gaps and overlaps when compared to the multi B-spline transforms without the sliding constraint. (paper)