| Radiotherapy is one of the three main treatment methods for cancer patients, the goal of radiotherapy is to maximize the therapeutic gain, which means to give high dose to the Planning Target Volume (PTV), while protect the surrounding normal tissues and Organs at Risk (OARs) from receiving higher or unnecessary dose at the same time. Intensity Modulated Radiotherapy (IMRT) is one of the most advanced radiotherapy techniques at present, and also the first choice in clinic currently. The principle of IMRT is to get conformal high dose distribution to target with non-uniform fluence map from the modulation of Multi-Leaf Collimator (MLC). The realization of IMRT technique calls for MLC in hardware and inverse optimization algorithm in software. The generation of IMRT treatment plan is as follows, when given a group of optimization parameters, the plan optimization problem was solved by inverse optimization algorithm to get the optimized fluence map, the second step is to segment the fluence map into MLC leaf sequences, finally get the treatment plan with ideal dose distribution. The deliver techniques of IMRT are various, including static IMRT and dynamic IMRT.Because of the variety of patient geometry, we will get different dosimetric end points for different patients. So the final ideal goal for treatment planning is to generate the optimal state plan for each patient, however, it is quite difficult to ensure the optimal state for each patient plan in the real clinic. Normally, a clinical IMRT plan is created guided by institutional protocols. These protocols define the plan dosimetric goals including PTV high dose coverage and OARs tolerance dose for each specific tumor site. In reality, medical physicist/dosimetrist follow these protocols to design the patient plans, as soon as they achieved these dosimetrical criteria, they don’t take extra time to improve the plan quality further, as a result, most of clinical treatment plans are stopped at these protocols. However, these plan dosimetric goals are obtained from statistic studies, sacrificing patient specificality. Hence, it is hardly to guarantee there is no more space to improve the PTV high dose coverage or spare OARs dose even the dosimetric goals have already been obtained during planning. Besides, the clinic treatment plan is obtained by medical physicist/dosimetrist performing multiple times trials and errors according to their own experience. The final quality and the efficiency is highly related to how experienced the planner is, how familiar with DVH objects setting affecting the corresponding dose distribution, how the final dose distribution should be according the patient geometry information, how hardness to achieve institutional protocol goals for each specific patient, and also the communication between planners and physicians, etc, all these mentioned factors can affect the final plan quality which is to be delivered to the patients. Unfortunately, there is no related plan quality control (QC) tools or modules in the clinic yet.Therefore, to address this clinical IMRT plan quality variation problem, ensure patients receiving the best radiation therapy, facilitate the communication between different institutions and the information sharing, assist or lead planners creating high quality plan before submitted, build scientific model, develop powerful tools for clinic use, is very meaningful and significant for both scientific research and clinic application.Plan quality control study is in a primary but popular stage, it can be overcome by observing the fundamental reasons as:1) the limitation of plan optimization algorithm; 2) the unknown of the optimization parameters. Accordingly, plan quality control can be conducted based on plan re-optimization and previous knowledge learning. The solution domain is limited to a partial Pareto surface for current optimization algorithms in commercial treatment planning systems, which are organ-weighting factor-based. When the optimization objective functions are modified to voxel weighting factor-based systems, the solution domain can be expanded to the entire Pareto surface, hence increasing the probability of finding a more optimal plan. Although studies have demonstrated the potential feasibility of applying voxel-weighting factor-based re-optimization algorithms to plan quality improvements, some other issues may need to be addressed for practical plan QC tools. Knowledge based planning is commonly applied to perform plan quality control and it can be conduct as the following procedures:patient geometry feature extraction, plan dosimetric feature extraction and correlation model establishment. The plan dosimetric features can be predicted by applying correlation model onto newly extracted geometry feature of a patient. Though certain prediction accuracy was achieved by previous studies, still, there is room for improvement. First, these tools adopt the entire DVH as the criteria for judging plan quality. Though DVH curve is commonly used for plan quality judgment, it contains information more than necessary for routine clinical usage. Planners normally pay more attention on dosimetric endpoints (DEs) than other dosimetric values unrelated to clinical goals. From this point, the model, directly adopting clinical goal representatives, such as DEs, would be more efficient and accurate. Second, the correlation model in some studies is simple and intuitive, uncertainty cannot be avoided with manual plan quality determination. Third, in most previously developed geometry-dosimetry models, OARs are considered independently, where dosimetric values of one OAR (or DVH) are the function of its own geometry only and are not affected by other OARs. However, the trade-off between multiple OARs dose-sparing cannot be avoided, indicating that dosimetric goals of one OAR dose sparing is affected by multi-OARs.By reviewing related papers, we did some research and achieved some results. These works including:First, based on the voxel-based optimization model, we successfully enhanced an automatic treatment re-planning system into an automatic IMRT treatment plan quality control tool. This IMRT treatment plan QC tool is an application of a voxel weighting factor based treatment re-optimization algorithm which is able to enlarge the plan optimization solution domain from part of Pareto surface to the whole Pareto surface, hence increasing the possibility to find the optimal result. By comparing the plan quality between under-assessment plan and system generated re-optimized plan, plan quality determination of under-assessment plan could be obtained. And the plan QC could be achieved by maintaining the high quality plan while replacing poor quality plan with system re-optimized plan.We developed a voxel-based re-optimization guided automatic treatment plan quality control tool for intensity modulated radiation therapy by adding extra modules into an automatic treatment re-planning system, including plan protocol importation, plan QC report generation, etc. By comparing the plan dosimetric endpoint values between under-assessment plan and SCORE re-optimized plan using a pairwise Wilcoxon p test, statistical significance of the difference between these plan pairs could be obtained to determine under-assessment plan quality. Plan would be identified as high quality when no significant difference observed, and poor quality with significant difference.To evaluate module functions and feasibility, plan QC determination efficacy and accuracy of the SCORE as a plan QC tool,25 clinically approved cervical patients IMRT plans and 5 poor quality IMRT plans which were generated on purpose were included. These 25 clinical approved cases were created with Eclipse system by UCSD Moores cancer center experienced dosimetrists with enough time and efforts, hence could be regarded as high quality plans. The SCORE QC function accuracy and feasibility could be evaluated by checking the consistence of QC tool plan quality determination and the known quality of the 25 clinical group and 5 poor quality group.Plan verification results show that, no significant difference could be observed between under-assessment plan and SCORE re-optimized plan dosimetric endpoint value arrays for the 25 clinical cervical cancer patients, indicating the under-assessment plan quality is acceptable by the extend QC tool quality determination. As to the other 5 poor quality cases, the extended QC tool presented significant differences on the given results, namely that the 5 cases were identified as poor quality plans with a large room for improvement. The plan quality judgment results given by developed QC tool and the actual plan quality are with high consistency, proving the feasibility and accuracy of the system as a QC tool.From the engineering point of view, this work is the first time realization of using plans re-optimization approach to perform plan quality control. Though many articles have already pointed out that the proposed planning optimization method was capable to conduct plan quality evaluation and control, but none of them realize it. In addition, the added QC related module is designed reasonable, smoothly linking to the original SCORE system, with generating plan QC report fast and accurately, making the whole plan QC procedure as a highly automated integration. Besides, less than 2 minutes could be achieved for each individual plan quality control, this high efficacy is allowing the plan QC function fully prepared for the next step of clinical use.We successfully developed a correlation model which is able to predict clinical relevant treatment plan dosimetric endpoints by considering multiple OARs geometry features. Given a new patient with its own geometry, the predicted achievable plan dosimetric endpoints could be used to compare with under-assessment plan and hence determine the treatment plan quality. By approving high quality plan and rejecting and perform re-planning poor quality plan, treatment plan quality could be controlled.We developed an experience-based treatment plan quality control, namely patient specific geometry predicting achievable plan dosimetric goals based automatic treatment plan quality control. The main idea is to first dig the potential related information from experience, then adopt appropriate machine learning methods to correlate patient geometric features with plan achievable dosimetric goals, hence establish a prediction model and mechanism, eventually product plan quality control. A new patient’s expecting plan dosimetric goals could be obtained by applying its new geometric feature into the prediction model. These predicting goals could act as the quality standard of the under-assessment plan, by comparing the extracted dosimetric endpoint values of current plan to the predicting goals, plan quality would be determined. Hence the plan quality could be controlled and improved by retaining high quality plans but rejecting and re-optimizing poor quality plans.Follow the aforementioned clue, we proposed a new geometry-dosimetry model where the geometry of each patient’s internal anatomy is used to predict optimally achievable values of dosimetric endpoints related to treatment outcome. The geometric parameters adopted in our model are sufficient to model the PTV and OAR geometric relationship, where an OAR volume is divided and binned along its distance to the PTV boundary, as the dDTH. The dDTH is the differential way of OVH which is proposed by Binbin Wu in 2009. Base on that, we simplified the relationship between dDTH and dosimetric endpoint values with a linear model set, and the coefficients can be retrieved through a linear regression method. To avoid the over-fitting issue of the linear regression approach, an L2-regularization term and a smoothing term are added to the model when performing the least-squares estimation. In order to get a set of reasonable coefficients with limited patient data, we adopt a cluster ensemble method:the bagging (bootstrap aggregating) algorithm to get rid of the sample outlier. Fifty clinically treated high quality prostate cancer patients VMAT plans with non-toxicity reported were chosen from our institute’s clinical database as our model training dataset. Another 20 VMAT plans from outside of the model training pool for model evaluationTraining sample fitting error shows the accuracy of correlation model between patient geometry and plan dosimetric endpoints well with small fitting error and standard deviation which is within 3%. According to the fitting error distribution, a prediction model based plan quality determination mechanism was proposed by a choosing 2 sigma confidence interval. Following the prediction model and plan quality determination mechanism, six out of the 20 evaluation plans were identified as suboptimal plans with further room to improve unsatisfied plan dosimetric endpoints. To prove the accuracy of the judgment, these 6 identified suboptimal plans were re-optimized and the re-optimized plans were recruited to undertake the plan QC process again. The updated evaluation results show that the re-optimized plans could achieve a better OARs dose sparing without sacrificing any PTV dose, besides, the model-predicted dosimetric endpoint values show an improved agreement with those extracted from the re-optimized plan. These results fully illustrate the accuracy and feasibility of the proposed prediction model.Compared with other experience-based plan quality control methods, our proposed prediction model and plan quality determination mechanism directly is targeted the most clinical relevant plan dosimetric endpoint by considering the effect from multiple OARs geometry features. The model coefficient gives an easy understanding of how the patient geometry affect plan dosimetric endpoint. Besides, the quantitative plan quality determination mechanism is innovative and the prediction model is accurate and effective. Furthermore, this prediction model is not only able to perform plan quality evaluation and control, ensuring patient’s best treatment, but also guide plan optimization during plan creation and hence act as a powerful tool for automatic planning.High quality clinical plan is the guarantee of radiotherapy technology towards more accurate, more efficient scientific research direction, is a key way to ensure that patients get the best treatment, before the treatment plan is delivered on patients, the effective quality evaluation and control is important for both scientific value and clinical significance. In this thesis, we focused on the popular research direction in this field, and obtained some preliminary research results, but we still need the further improvement. |
PDF Full Text Request