| Radiation therapy is one of the main three means to treat malignant tumor,70% of cancer were treated by radiotherapy in China. Over the years, stereotactic radiosurgery (SRS), three-dimensional conformal radiotherapy (3D-CRT) and intensity-modulated radiation (IMRT) and image-guided radiation therapy (IGRT) have been developed and applied in clinical, which brought good news for cancer patients. However, organ movement in thorax and abdomen mainly caused by breathing during radiotherapy, may make the target tumors escape and normal tissue around the tumor into the planning target volume(PTV), thus greatly affect the efficiency of intensity-modulated conformal radiation therapy, and increases the probability of complications. The current means of conventional radiotherapy is to amplify the planning target volume, namely including the movement amplitude into the PTV, in order to ensure clinical target volume (CTV) is always included in the PTV during the process of radiotherapy. However, this measures that it will make more normal tissue around the tumor into the beam, and the total radiation dose increases exponentially with target diameter increases. Therefore, how to compensate tumor motion mainly caused by the breathing has become one of the biggest and inevitable problems to realize precise radiotherapy for tumors in thorax and abdomen.At present, a variety of methods had been proposed to solve the problem of the tumor motion caused by respiration in radiotherapy, there are mainly including breath-hold technology or abdominal compression, respiratory gating technology,4-D radiation therapy method and real-time tracking technology. (1) Breath-hold technology or pressure given methods to control technology patients’ respiratory actively or passively to reduce tumor motion; the methods is simple but with poor tolerance and low precision.(2) Respiratory gating technology to make the X-ray beam exposure is synchronized with respiratory cycle in a particular phase, which can reduce the tumor motion during the exposure time window; the disadvantage is that linear accelerator opens the beam only in a particular phase, which causes the reduce of the working cycle of linear accelerator, and makes radiation time more longer.(3) The four-dimensional radiation technology, which has considered the anatomy changes over time through the whole radiation stage including image positioning, design of treatment planning and implementation on the basis of three dimensional radiation therapy; the disadvantages are the long image acquisition time and motion artifacts in the vertical direction, patients also need to have breathing exercises to ensure keeping consistent during the whole process.(4) Real-time tumor tracking technology, the advantage is that patients can breathe normally, and can following tumor motion in real time so that to adjust the position of the beam or treatment bed to ensure relative static between the beam and the target tumor. Currently there are three most widely used methods:1,To track implanted metal markers near the target tumor via X-ray imaging, which can accurately obtain the movement information of internal target tumor, the disadvantage is that the method is invasive, and the higher the imaging frequency, the more radiation dose patients received, but the low frequency will lead to low accuracy.2, the optical positioning system to get the patient’s body surface movement information, the advantages are convenient operation and completely noninvasive, the disadvantage is that the relationship between the tumor and the surrogate is not constant, so it is very difficult to achieve the accurate tracking of target tumor in the body only by in the surrogate.3, The most feasibility and valuable method is indirect tracking mode, which makes full use of the advantages of the above two methods, it is also using the correlation of the target tumor and the surrogate, but the difference is that the hybrid method at first acquires some internal and external data synchronously, then calculates the tumor’s position using the surrogates’ data collected. At present the most clinically advanced technology-Cyberknife’s synchronous respiratory tracking system is based on the idea.Through careful study of the principle of the Cyberknife system, this paper is dedicated to achieve a dynamic real-time tumor tracking technology based on the surrogate markers. The theoretical foundation of the project is the good correlation between external respiration signal and the tumor motion, the experimental results further show that the correlation between the surface markers and tumor motion is with an average of 0.77, and in the range between 0.41 and 0.97. As can be seen, the tumor-surrogate correlation is not a simple linear relation, so it needs to be modeling with more accurate function. In addition, because the tumor motion is not the same as in the inspiratory and expiratory, which is called hysteresis phenomenon, it is also a respiratory characteristics need to capture. Thereby, in order to achieve indirectly tracking system, the key premise is to found a correlation model can accurately express the tumor-surrogate correlation and fully capture the movement characteristics of breathing. In addition, whether the signal acquisition and processing, or invoke the new radiation parameters and hardware response such as multi leaf collimator, treatment bed, all require a certain amount of time, that is to say system delay is inevitable from the signal acquisition and processing to the adjustment of the beam, so prediction algorithm is essential to compensate system delay.Currently, plenty of literatures had been Researching correlation model and prediction algorithm, but these methods have some shortcomings, and thus become a bottleneck restricting the development of real-time tracking radiation technology. This paper studies on the proposed new algorithm to construct a more accurate and robust correlation model and prediction algorithm, so as to build a theoretical foundation for the realization of the indirect tracking system. The paper has proposed a memory-based learning method to construct the prediction model. And locally weighted linear regression is proposed to build correlation model in this paper, the results show the advantages of strong robustness, high accuracy and in real-time. They firstly store data in memory as the training set, and then find the relevant data from memory to answer particular queries. The nearer data point is designed the higher relevance (or weight). By relatively simple models rather than a global model, the method can accurately capture extremely complex and nonlinear relationship. In addition, due to the local nature of the weighted function, the method is robust to reject the outliers in the training set. Furthermore, the method training and adapting to the new data are almost in real time. The above features making memory-based learning method has fairly good accuracy and robustness.Considering a major problem in modeling the respiratory motion-hysteresis, the paper adopts a solution of "state augmentation", where the input vector is augmented by the current data with one or several time-lagged data with a certain time lag, so that to capture the recent breathing dynamics. In general, accuracy of the algorithm increases with expansion of the training set. However, oversize training set can lead to slow running speed. Furthermore, the respiratory motion irregularly changes over time, and near data point has better relevance than distant data. Therefore, in order to considering this feature into the algorithm, we use the "sliding window" to dynamically update the training set, that is first form suitable training set size, when the new data is collected, the old data will be replaced by the new data based on the "FIFO" rule. It can not only keep the training set in suitable size but also always contain the latest data, and improve the accuracy of the algorithm. The core concept of memory-based learning method is to use relevant data to calculate the output value. It makes full use of each data point in the training set in order to calculate the output value. Because the distance function depends on the new input, each training data need to be stored in memory to generate the appropriate output values. Local weighted training is used for the data, which requests local model should be fitting well on neighbor data points, and less attention to distant data points.the Gaussian kernel function is used to measure relevance, the weighted factor is applied to the each data point in the train set (input vectors).The optimal values of model parameters are calculated by local weighted regression, but it leads to memory-based learning will face a problem of ’morbid matrix’ of the regression analysis method in some cases, thus greatly increase the mean square error of regression coefficient, and influence the accuracy and robustness of the algorithm. The paper innovatively proposes ridge regression to eliminate the influence of’morbid matrix’ and to improve the performance of the algorithm. Ridge regression, in a nutshell, is to add a suitable nonnegative factor into the main diagonal elements of’morbid matrix’, which would makes a slight deviation estimates of regression coefficient but can significantly improve the stability of the predictor. In fact, a variety of regression analysis methods are widely applied to construct the correlation model or prediction model, and they often meet with abnormal state, but generally adopt the solution of expanding size of the training set,but it usually cause low speed, so difficult to meet the real-time requirements. Therefore, the ridge regression method proposed in the paper also has very good reference significance for other regression analysis.In order to prove the performance of memory-based learning to predict respiratory motion, the experiment acquires the data from 10 volunteers’ body surface markers based on POLARIS optical positioning system (the average motion amplitude is about 20 mm), then to forecast respectively using linear regression, conventional memory-based learning, and memory-based learning improved by ridge regression. The results show that the algorithm we proposed has a high prediction precision (the average of mean absolute error is about 0.3 mm) even if under the condition of long time delay(1s), and each prediction only takes about 1 ms so it can achieve to accurately predict respiratory motion in real-time. The performance is not only better than linear regression, and ridge regression can eliminate the abnormal state to greatly enhance performance of the basic memory-based learning. Furthermore, the surrogate data based on accuTrack250 system is used to prove effective accuracy and robustness of prediction model via memory-base learning.Locally weighted linear regression improved by ridge regression is proposed to build correlation model in this paper, the results show the advantages of strong robustness, high accuracy and in real-time. Experiment uses 7 sets of synchronous in-out data, where the internal maker’s motion data acquired by 3D ultrasound for liver blood vessels, and the external makers’ motion data acquired by optical positioned system. The results show that the error much less than the conventional models, can satisfy the requirement of real-time tracking technology. Along with the further research, the selection of model parameters will make better choices, and the accuracy of the model can also be improved. Finally, the prediction model and the correlation model have been combined in our research, Experiment uses 7 sets of synchronous in-out data above-mentioned, and the results preliminarily verify accuracy of the hybrid model. At the end of the paper, we make a summary and outlook for our research work, which illustrates some shortcomings in the research, and the future works are introduced. |
PDF Full Text Request