Research Of Re-optimization For Adaptive Radiotherapy And Automated Treatment Planning

Radiation therapy is one of the most important techniques for cancer treatments, and above50%of the patients diagnosed with canceris treated with radiation therapy, so it is valuable to focus on furthering the advancements of radiation therapy to provide improved treatment for cancer patients. The goal of radiation therapy is to deliver high dose to the malignant cells (i.e. the tumor), while sparing the surrounding healthy tissues as much as possible.Currently, There-Dimensional Conformal Radiation Therapy(3DCRT) and Intensity Modulated Radiation Therapy(IMRT) are the most commonly used technologies in clinic.3DCRT aims radiation to the tumor from multiple directions. The beam intensity of each treatment field is uniform and beam aperture is shaped to match the projection of the target.Unlike3DCRT, IMRT beam intensity varies across the treatment field to deliver highly conformal dose distribution to target while sparing surrounding organs and normal tissues. IMRT is a valuable technique since it is able to produce steeper dose gradientsat target-normal tissue interfaces as compared to3DCRT.The ability to generate the steep dose gradients makes IMRT can deliver accurate dose to the complex static target, but also makes it prone to setup errors and anatomic changes, which may pose risks for a target missing-Recently, CBCT has been used in Image Guided Radiotherapy(IGRT), which can correct for setup uncertainties and improve the treatment accuracy to a certain extent. However, the uncertainty due to organ deformation still remains. As the approved treatment plan which is designed before treatment will be used to treat a patient for the entire course of many fractions, but the geometry of tumor and normal anatomy change significantly during treatment course, which will definitely deteriorate the pre-designed treatment plan quality. In order to solve this problem, a lot of attention have been attracted to Adaptive Radiation Therapy(ART). ART with the basic idea of imaging the patient’s changed geometry during the treatment course and modifying the treatment plan accordingly, has been proposed by Yan.It is a novel approach to reduce normal tissue toxicity and/or improve tumor control.There are two ways to implement ART. One way is based on offline re-planning, and the other is based on online re-planning. For both methods, IMRT re-planning is an important component. As radiation therapy has a dual objectives:irradiate the tumor and spare the normal tissues, IMRT planning is a multi-criteria optimization(MCO) problem. The most prevalent methods to solve a MCO problem is by transforming it into a single objective problem using a specific set of weighting factors for each structure-specific planning objective. As the weighting factors have no clinical meaning and the choice quite arbitrary,the search for suitable weights often requires many re-optimizations. For the re-planning process, manual trial-and-error approach to fine-tune planning parameters is time-consuming and is usually considered unpractical, especially for online ART,where the patient is still lying on the treatment couch and a new plan is needed immediately.It is desirable to automate this step to yield a plan of acceptable quality with minimal interventions.Moreover, for offline ART, although the requirementfor a very fast treatment planning process is not that strong, it is still preferable to improve the planning efficiency.Developing an efficient and effective re-planning algorithm is an important step towards the clinical realization of ART.As mentioned above, although IMRT is a standard technique in radiation therapy, there are still many problems with current treatment planning method. For example, current planning is conducted in a low efficiency but high cost way. The whole process may take several hours, even a week, for clinicians to generate a clinically acceptable plan for one patient by manually tuning the planning parameters iteratively in a trial-and-error way. More importantly, under this kind of process, plan quality heavily depends on the experience of the planner and the amount of time spent on the plan Large variation in plan quality and time for planning have been reported in multi-institution studies, which makes it difficult for crossing institution collaboration, experience sharing, data sharing and multi-institution clinical trials. So it is desirable to automate this step to yield a plan of acceptable quality with minimal interventions, which will solve the problems with current planning method, including time-consuming, resource-requiring and having large variation in plan quality.In order to meet the requirement of a very fast treatment planning process and to solve the problems with current planning method, we did a research on re-planning method for ART and automated treatment planning methods, based on Super Computer Online Re-planning Environment(SCORE), a re-planning system developed at the University of California, San Diego. This research mainly contains:(1) A comprehensive review was made on the current status of the automated treatment planning methods and re-planning methods. Then we described Fluence map optimization method, which is the most commonly used in clinic. And based on this, summarized the dose calculation model and optimization methods used in our work. Abriefintroduction about the re-planning platform SCORE was also given, which is essential for our research.(2) To meet the requirement of ART, we proposed an automatic re-planning algorithm to generate a plan with similar, or possibly better, DVH curves compared with the clinically deliveredoriginal plan. Specifically, our algorithm uses a voxel-based model and iterates the following two loops. An inner loop is the traditional fluence map optimization, in which we optimize a quadratic objective function penalizing the deviation of the dose received by each voxel from its prescribed or threshold dose with a set of fixed voxel weighting factors. In outer loop, the voxel weighting factors in the objective functionare adjusted according to the deviation of the current DVH curves from those in the original plan. The process is repeated until the DVH curves are acceptable or maximum iteration step is reached. The whole algorithm is implemented on GPU for high efficiency. The feasibility of our algorithm has been demonstrated with three head-and-neck cancer IMRT cases, each having an initial planning CT scan and another treatment CT scan acquired in the middle of treatment course. Compared with the DVH curves in the original plan,the DVH curves in the resulting plan using our algorithm with30iterations arebetter for almost all structures. The re-optimization process takes about30secondsusing our in-house optimization engine, making it possible for clinical usage.Compared with other re-planning methods, the proposed method has the following advantages:Firstly, we are using a voxel-based model, where a single weighting factor is assigned to each voxel in the objective function. A lot of works have shown that a voxel model can explore a much larger Pareto surface than an organ model, likely leading to a plan with better trade-off. We also proved that, while the objective function selection is a big concern for the organ-based model, it is no longer an issue for the voxel-based model and the entire large Pareto surface could be explored by using a simple quadratic objective function. Secondly,the proposed method using the DVH in the original plan as guidance have two advantages. First of all, DVH curves are useful tools for plan evaluation that have direct clinical significance. Since the DVH curves from the original plan imply information of approved dose-volume constraints and represent clinician-approved trade-offs among different structures for this particular patient, we can use them as the desired DVHs to guide the re-optimization for the patient’s new geometry to ensure the clinical acceptability of the new plan. Here, we assume that the original DVHs contain the clinician-approved organ trade-off information and clinically desirable for the patients’ new geometry. Second, using DVH information as guidance will give more flexibility for optimization than using dose distribution. It is more likely to reproduce a given DVH than to reproduce a given dose distribution, since the DVH is insensitive to the permutation of voxel doses within each organ.Moreover, the proposed method can generate a fully optimized IMRT plan as opposed to heuristically modifying the original one. It uses the DVH in the original plan as guidance to automatically adjust the weighting factor and is implemented on GPU for high efficiency, which will meet the efficiency requirement of ART.(3) In order to solve the problems with current treatment planning method, we proposed an automated treatment planning process by suing a GPU-based automatic planning engine and a library of previously delivered treatment plans. First of all, a library of100prostate IMRT plans previously delivered at MooresCancer Center has been assembled to build a planning library. Then GPU-based platform SCORE can be automatically launched for many times using each plan from the library as reference to generate a set of deliverable plans with different tradeoffs.After a set of deliverable plans were generated for a new patient using our GPU-based platform. A GUI was developed to allow clinicians to navigate through the deliverable plans at high efficiency to select the best plan using institutional criteria for treatment plan quality.To evaluate the effectiveness of the proposed method, leave-one-out cross validation was performed whereby each of the100patients was selected as a "new" patient and the other99patients were treated as reference patients.Finally, compared with the original clinical plan, each of the final plans was evaluated by an experienced physician using DVH curves and the specific plan quality metrics. And then SCORE plans were classified into three categories:better than, equal to and worse than, the original plan. The numbers of final SCORE plans for the100patients in the three categories were44,45and11respectively. And by checking the specific DVH constraints for each patient, we found that almost all the SCORE plans had a significantly more homogeneous dose to the planning target volume and a significantly lower maximum dose for femoral heads while keeping the same quality for other OARs. So the quality of the majority SCORE plans was better than or equivalent to that of the original plans in terms of DVHs either based on the specific plan quality metrics or the physician’s experience. Besides providing high-quality IMRT treatment plans, the proposed method may dramatically reduce treatment planning time and effort. For one prostate IMRT patient, the time to generate99new plans by automatically running SCORE without any manual intervention was consistently between40and100minutes, with an overall average of82minutesbased on our leave-one-out experiment. Then it takes less than1minute to select a desired plan from the99plans through the plan selection GUI.In summary, this research addressed two important challenges related to treatment planning. Firstly, we have successfully developed an automatic re-optimization algorithm, whichis an important step towards the clinical realization of ART.In the proposed re-optimization algorithm, the automated weighting factors adjustment process make it possible to yield a plan of acceptable quality with minimal interventions. And based on the re-optimization method, we proposed an automatic treatment planning procedure, which can be a solution to the current manual treatment planning process requiring significant human effort, planner experience, and clinicians. Moreover, both of the methods are implemented on GPU, so a high computational efficiency has been achieved, making it possible for clinical usage.