Radiation therapy is one of the main treatment methods for cancer. Machine learning can be used in all aspects of clinical practice in radiation therapy, including clinical decision support, automatic segmentation of target volumes, prediction of treatment efficacy and side effects. Despite the challenges of lacking structured data and poor interpretability of models, the application of machine learning in radiotherapy will become increasingly profound and extensive. This review contains three aspects: introduction of machine learning, the clinical application of machine learning in radiotherapy, challenges and solutions.

Objective To develop whole breast irradiation with volumetric-modulated arc therapy (VMAT)-based hybrid intensity-modulated radiotherapy (IMRT) after breast conserving surgery for breast cancer,and to evaluate its value in clinical application.Methods Ten patients with breast cancer undergoing breast conserving surgery were enrolled.Two radiotherapy plans were designed based on hybrid fixed-beam IMRT/IMRT technique and hybrid VMAT/IMRT technique,respectively.One group received whole breast irradiation with a prescribed dose of 50 Gy in 25 fractions.The other group received whole breast irradiation with a prescribed dose of 50 Gy in 25 fractions,as well as simultaneous integrated boost to the tumor bed with 60 Gy in 25 fractions.The dosimetric parameters and delivery efficiency were compared between the two plans using paired t test.Results For patients treated with whole breast irradiation alone,there were no significant differences in conformity index and homogeneity index of target volume between the two plans (P=0.866,0.056);however,compared with the hybrid fixed-beam IMRT/IMRT technique,the hybrid VMAT/IMRT technique significantly increased the doses to organs at risk (OARs) and the number of monitor units (MUs) in the intensity-modulated field (P=0.000-0.050,P=0.002).For patients treated with whole breast irradiation with a simultaneous integrated boost to the tumor bed;however,the hybrid VMAT/IMRT technique significantly reduced the doses to the lung and spinal cord,number of MUs in intensity-modulated field,and delivery time compared with the hybrid fixed-beam IMRT/IMRT technique (P=0.004,0.001,0.000,0.000).Conclusions For patients treated with whole breast irradiation with a simultaneous integrated boost to the tumor bed,the hybrid VMAT/IMRT technique is highly promising for clinical application due to better OAR sparing and higher delivery efficiency.

Magnetic resonance imaging (MRI) simulator (MRI-Sim) can provide superior images for radiotherapy.Due to the complexity of MRI technology and the safety problem caused by strong magnetic field, the acquisition and implementation of MRI simulation is more complicated than CT simulation.In order to ensure the introduction of MRI-Sim, this paper reviews the selection, installation, and acceptance test of MRI-Sim, including the selection of host and auxiliary equipment, installation site preparation, and safety precautions,as well as MRI-Sim acceptance test and commissioning.

Objective To evaluate the plan quality of volumetric modulated arc therapy (VMAT) in single brain metastasis compared with the stereotactic arc therapy (S_ARC).Methods 31 patients were replanned using VMAT and S_ARC technique.Prescription dose is 40 Gy delivered in 4 fractions covering at least 95％ of the target volume while keeping minimum doses to the volume of normal brain tissue.The plans were assessed and compared using the conformity indexes (CI),gradient indexes (GI),the mean dose of normal brain tissue,the volumes of normal brain tissue receiving 4 Gy doses,the number of monitor unit and treatment times.A paired t test or non-parametric Wilcoxon signed rank test was performed to analyze the difference between these two plans.Results VMAT plans increased dose conformity,but not dose gradient,compared with S_ARC plans.The median dose conformity index values were 0.815,0.818,and 0.779 (P =0.000,0.000),and the median dose gradient score index values were 5.865,5.706,and 3.133(P =0.000,0.000,0.000),in single arc,double arc VMAT plans and S_ARC plans,respectively.The mean dose of normal brain tissue was higher in VMAT plans.And the volume of the normal brain tissue receiving doses of 4 Gy in VMAT plans was significantly larger than the volume in S_ARC plans.The VMAT plans got higher MU number (P =0.000,0.000).And the mean treatment times were (2.7 ± 0.1) min,(2.8 ± 0.1) min,and (7.6 ± 0.2) min in single arc,double arc VMAT plans and S_ ARC plans,respectively (P =0.000,0.000).Conclusions The dosimetry of VMAT plan can meet the requirements of clinical,and is recommended in the treatment of big volume single brain metastasis or single brain metastasis in cerebellum.

Objective To evaluate the effects of four types of variance reduction techniques ( ring counter grid, high electron cutoff energy, termination of electron tracking in some structures, and emission direction?biased sampling of source) on the efficiency and accuracy of Monte Carlo simulation for the single source channel of the Leksell Gamma Knife. Methods The single source channel of the Leksell Gamma Knife was modeled using Monte Carlo software MCNP . Four types of variance reduction techniques were used to simulate the dose distribution in the water?like phantom. The computation efficiency and simulation result were compared between the four techniques. Results All techniques substantially improved the computation efficiency and had little effect on the accuracy of the simulation ( relative error less than 2. 5%) . However, if the electron cutoff energy was above 50 keV, the simulation became quite inaccurate due to neglect of the scattering of high?energy electrons and their dosimetric contribution to the penumbra. When the scattering of high?energy electrons and their dosimetric contribution to the penumbra were ignored, the dose to the Profile platform was overestimated and the dose to the penumbra was underestimated. Conclusions Rational use of variance reduction techniques can substantially improve the efficiency of Monte Carlo simulation for the single source channel of the Leksell Gamma Knife. However, the impact of variance reduction techniques on the accuracy of the simulation should be carefully evaluated.

Objective The aim of this study is to evaluate the feasibility of using orthogonal kilo voltage fluoroscopic imaging(OKVFI)for setup correction in image guided radiotherapy of the liver.Methods After positioned the patients with liver cancer implanted with silver rings on the accelerator equipped with kilo voltage X-ray volume imaging(XVI),averaged OKVFI and cone beam CT(CBCT) volumetric images were acquired.A total of 90 datasets of averaged OKVFI and 90 datasets of volumetric images for 10 patients were obtained.The couch shifts obtained by the matching between OKVFI and digitally reconstructed radiograph were compared tu those achieved by the registration between CBCT and 4D reference average CT.On the comparison of the two different matching metheds.the Pearson coefficient was used to analyzed the correlation and Bland-Altman analysis to discern the consistence.Results The Pearson coefficient of correlation for the patient position shifts were R2=0.821.0.771 and 0.909 in the left-right (LR),anterior-posterior(AP)and superior-inferior(SI)directions respectively.95% CI were-2.30 -1.53(LR),-2.06-3.01(AP)and-2.69-1.53(SI)respectively.Within a±3 mm tolerance were 97.78%.95.56%and 96.67%respectively.Conclusions OKVFI has hish correlation and consistence with CBCT image on the setup correction.It is feasible to implement position correction with OKVFI in clinic practice.

The frequency-domain magnetic resonance spectroscopy (MRS) is achieved by the Fast Fourier Transform (FFT) of the time-domain signals. Usually we are only interested in the portion lying in a frequency band of the whole spectrum. A method based on the singular value decomposition (SVD) and frequency-selection is presented in this article. The method quantifies the spectrum lying in the interested frequency band and reduces the interference of the parts lying out of the band in a computationally efficient way. Comparative experiments with the standard time-domain SVD method indicate that the method introduced in this article is accurate and timesaving in practical situations.
Algorithms ; Brain ; metabolism ; Computer Simulation ; Fourier Analysis ; Humans ; Magnetic Resonance Spectroscopy ; methods ; Signal Processing, Computer-Assisted

Objective:To investigate a time series deep learning model for respiratory motion prediction.Methods:Eighty pieces of respiratory motion data from lung cancer patients were used in this study. They were divided into a training set and a test set at a ratio of 8∶2. The Informer deep learning network was employed to predict the respiratory motions with a latency of about 600 ms. The model performance was evaluated based on normalized root mean square errors (nRMSEs) and relative root mean square errors (rRMSEs).Results:The Informer model outperformed the conventional multilayer perceptron (MLP) and long short-term memory (LSTM) models. The Informer model yielded an average nRMSE and rRMSE of 0.270 and 0.365, respectively, at a prediction time of 423 ms, and 0.380 and 0.379, respectively, at a prediction time of 615 ms.Conclusions:The Informer model performs well in the case of a longer prediction time and has potential application value for improving the effects of the real-time tracking technology.

Objective:To investigate the pseudo-CT generation from cone beam CT (CBCT) by a deep learning method for the clinical need of adaptive radiotherapy.Methods:CBCT data from 74 prostate cancer patients collected by Varian On-Board Imager and their simulated positioning CT images were used for this study. The deformable registration was implemented by MIM software. And the data were randomly divided into the training set ( n=59) and test set ( n=15). U-net, Pix2PixGAN and CycleGAN were employed to learn the mapping from CBCT to simulated positioning CT. The evaluation indexes included mean absolute error (MAE), structural similarity index (SSIM) and peak signal to noise ratio (PSNR), with the deformed CT chosen as the reference. In addition, the quality of image was analyzed separately, including soft tissue resolution, image noise and artifacts, etc. Results:The MAE of images generated by U-net, Pix2PixGAN and CycleGAN were (29.4±16.1) HU, (37.1±14.4) HU and (34.3±17.3) HU, respectively. In terms of image quality, the images generated by U-net and Pix2PixGAN had excessive blur, resulting in image distortion; while the images generated by CycleGAN retained the CBCT image structure and improved the image quality.Conclusion:CycleGAN is able to effectively improve the quality of CBCT images, and has potential to be used in adaptive radiotherapy.

Objective: To obtain the high-resolution dose distribution for the single source channel of Leksell 4C gamma knife. Methods: A parallel computing platform based on the Message Passing Interface (MPI) and Monte Carlo Code MCNPX was established. The ring-shaped detector and two pre-validated variance reduction techniques (emission direction-biased sampling of source and termination of electron tracking in partial structures) were adopted to derive the high-resolution dose distribution for the single source channel of Leksell 4C gamma knife. The effect of cut-off energy for both photon and electron on the accuracy of simulation outcomes was evaluated and statistically compared. Results: Compared with previous findings, the spatial resolution of the dose distribution for the single source channel obtained in this study was higher (radial resolution=0.1 mm) with less statistical error (<1%). The calculation time was acceptable (approximately 24 h). For the 4-, 8-, 14-and 18-mm variable collimators, the penumbra and full width at the half maximum (FWHM) for single side were 1.0, 1.1, 1.2, 1.3 mm and 2.2, 4.3, 7.3, 9.3 mm, respectively, which were consistent with previous studies. The difference of the simulation results was extremely small between different cut-off energy for photon (1 keV vs.10 keV). However, the simulation results significantly differ between 1 and 521 keV electronic cut-off energy. Conclusions The MCNPX parallel computing platform based on the MPI environment can be utilized to derive highly accurate dose distribution with high resolution in acceptable calculation time. The cut-off energy of the photon and electron should be cautiously set up during simulation.