Browsing by Author "Wu, Qingrong (Jackie)"

Purpose: To develop an automatic collimator setting optimization algorithm to improve dosimetric quality of pancreas Volumetric Modulated Arc Therapy (VMAT) plans for Stereotactic Body Radiation Therapy (SBRT).

Methods: Fifty-five pancreas SBRT cases were retrospectively studied. Different from the conventional practice of initializing collimator settings manually, the proposed algorithm simultaneously optimizes the collimator angles and jaw positions which are customized to the patient geometry. This algorithm includes two key steps: an iterative optimization algorithm via simulated annealing that generates a set of collimator settings candidates, and a scoring system that choose the final collimator settings based on organs-at-risk (OARs) sparing criteria and dose prescription. The scoring system penalizes 3 factors: 1) jaw opening ratio on Y direction to X direction; 2) unmodulated MLC area within the jaw aperture in a dynamic MLC sequence; 3) OAR shielding capability by MLC with MLC aperture control constraints. For validation, the other 16 pancreas SBRT cases were analyzed. Two dual-arc plans were generated for each validation case, an optimized plan (Planopt) and a conventional plan (Planconv). Each plan was generated by a same set of auxiliary planning structures and dose-volume-histogram (DVH) constraints in inverse optimization. Dosimetric results were analyzed and compared. All results were tested by Wilcoxon signed-rank tests.

Results: Both plan groups had no statistical differences in target dose coverage V95% (p=0.84) and Root Conformity Index (p=0.30). Mean doses of OARs were improved or comparable. In comparison with Planconv, Planopt reduced maximum dose (D0.03cc) to stomach (-49.5cGy, p=0.03), duodenum (-63.5cGy, p<0.01), and bowel (-62.5cGy, p=0.01). Planopt also showed lower modulation complexity score (p=0.02), which implies its higher modulation complexity of the dynamic MLC sequence.

Conclusions: The proposed collimator settings optimization algorithm successfully improved dosimetric performance for dual-arc VMAT plans in pancreas SBRT. The proposed algorithm was demonstrated with great clinical feasibility and readiness.

Introduction: VMAT treatment planning requires time-consuming DVH-based inverse optimization process, which impedes its application in time-sensitive situations. This work aims to develop a deep-learning based algorithm, Automatic VMAT Planning via MLC Dynamic Sequence Prediction (AVP-DSP), for rapid prostate VMAT treatment planning.

Methods: AVP-DSP utilizes a series of 2D projections of a patient’s dose prediction and contour structures to generate a single 360º dynamic MLC sequence in a VMAT plan. The backbone of AVP-DSP is a novel U-net implementation which has a 4-resolution-step analysis path and a 4-resolution-step synthesis path. AVP-DSP was developed based on 131 previous prostate patients who received simultaneously-integrated-boost (SIB) treatment (58.8Gy/70Gy to PTV58.8/PTV70 in 28fx). All patients were planned by a 360º single-arc VMAT technique using an in-house intelligent planning tool in a commercial treatment planning system (TPS). 120 plans were used in AVP-DSP training/validation, and 11 plans were used as independent tests. Key dosimetric metrics achieved by AVP-DSP were compared against the ones planned by the commercial TPS.

Results: After dose normalization (PTV70 V70Gy=95%), all 11 AVP-DSP test plans met institutional clinic guidelines of dose distribution outside PTV. Bladder (V70Gy=6.8±3.6cc, V40Gy=19.4±9.2%) and rectum (V70Gy=2.8±1.8cc, V40Gy=26.3±5.9%) results in AVP-DSP plans were comparable with the commercial TPS plan results (bladder V70Gy=4.1±2.0cc, V40Gy=17.7±8.9%; rectum V70Gy=1.4±0.7cc, V40Gy=24.0±5.0%). 3D max dose results in AVP-DSP plans(D1cc=118.9±4.1%) were higher than the commercial TPS plans results(D1cc=106.7±0.8%). On average, AVP-DSP used 30 seconds for a plan generation in contrast to the current clinical practice (>20 minutes).

Conclusion: Results suggest that AVP-DSP can generate a prostate VMAT plan with clinically-acceptable dosimetric quality. With its high efficiency, AVP-DSP may hold great potentials of real-time planning application after further validation.

Introduction: Several studies have realized fluence-map-prediction-based DL IMRT planning algorithms. However, DL-based VMAT planning remains unsolved. A main difficult in DL-based VMAT planning is how to generate leaf sequences from the predicted radiation intensity maps. Leaf sequences are required for a large number of control points and meet physical restrictions of MLC. A previous study1 reported a DL algorithm to generate 64-beam IMRT plans to approximate VMAT plans with certain dose distributions as input. As a step forward, another study2 reported a DL algorithm to generate one-arc VMAT plans from patient anatomy. This study generated MLC leaf sequence from thresholded predicted intensity maps for one-arc VMAT plans. Based on this study, we developed an algorithm to convert DL-predicted intensity maps to multi-segment VMAT plans to improve the performance of one-arc plans.

Methods: Our deep learning model utilizes a series of 2D projections of a patient’s dose prediction and contour structures to generate a multi-arc 360º dynamic MLC sequence in a VMAT plan. The backbone of this model is a novel U-net implementation which has a 4-resolution-step analysis path and a 4-resolution-step synthesis path. In the pretrained DL model, a total of 130 patients were involved, with 120 patients in the training and 11 patients in testing groups, respectively. These patients were prescribed with 70Gy/58.8Gy to the primary/boost PTVs in 28 fractions in a simulated integrated boost (SIB) regime. In this study, 7-8 arcs with the same collimator angle are used to simulate the predicted intensity maps. The predicted intensity maps are separated into 7-8 segments along the collimator angle. Hence, the arcs could separately simulate predicted intensity maps with independent weight factors. This separation also potentially allows MLC leaves to simulate more dose gradient in the predicted intensity mapsResults: After dose normalization (PTV70 V70Gy=95%), all 11 multi-segment test plans met institutional clinic guidelines of dose distribution outside PTV. Bladder (V70Gy=5.3±3.3cc, V40Gy=16.1±8.6%) and rectum (V70Gy=4.5±2.3cc, V40Gy=33.4±8.1%) results in multi-segment plans were comparable with the commercial TPS plan results. 3D max dose results in AVP-DSP plans(D1cc=112.6±1.9%) were higher than the commercial TPS plans results(D1cc=106.7±0.8%). On average, AVP-DSP used 600 seconds for a plan generation in contrast to the current clinical practice (>20 minutes).

Conclusion: Results suggest that multi-segment plans can generate a prostate VMAT plan with clinically-acceptable dosimetric quality. the proposed multi-segment plan generation algorithm has the capability to achieve higher modulation and lower maximum dose. With its high efficiency, multi-segment may hold great potentials of real-time planning application after further validation.

In external beam radiation therapy (EBRT) for cancer patients, it is highly desirable to completely eradicate the cancerous cells for the purpose of improving the patient’s quality of life and increasing the patient’s likelihood of survival. However, there can be significant side effects when large regions of healthy cells are irradiated during EBRT, particularly for organs-at-risk (OARs). Due to the juxtaposition of the cancerous and non-cancerous tissue, trade-offs need to be made between target coverage and OAR sparing during treatment planning. For this reason, the treatment planning process can be posed as a multi-criterial optimization (MCO) problem, which has previously been studied extensively with several exact solutions existing specifically for radiation therapy. Typical MCO implementations for EBRT involve creating, optimizing, and calculating many treatment plans to infer the set of feasible best radiation doses, or the Pareto surface. However, each optimization and calculation can take 10-30 minutes per plan. As a result, generating enough plans to attain an accurate representation of the Pareto surface can be very time-consuming, particularly in higher-dimensions with many possible trade-offs.

The purpose of this study is to streamline the MCO workflow by using a machine-learning model to quickly predict the Pareto surface plan doses, rather than exactly computing them. The primary focus of this study focuses on the development and analysis of the dose prediction model. The secondary focus of this study is to develop new metrics for analyzing the similarity between different Pareto surface interpolations. The tertiary focus of this study is to investigate the feasibility of deliberately irradiating the epidural space in spine stereotactic radiosurgery (SRS), as well as estimate its potential effect on preventing tumor recurrence.

For the primary focus of this study, the model’s architecture proceeds as follows. The model begins by creating an initial dose distribution via an inverse fit of inter-slice and intra-slice PTV distance maps on a voxel-wise basis. The model proceeds by extracting three sets of transverse patches from all structure maps and the initialized dose map at each voxel. The model then uses the patch vectors as inputs for a neural network which updates and refines the dose initialization to achieve a final dose prediction. The primary motivation behind our model is to use our understanding of the general shape of dose distributions to remove much of the nonlinearity of the dose prediction problem, decreasing the difficulty of subsequent network predictions. Our model is able to take the optimization priorities into account during dose prediction and infer feasible dose distributions across a range of optimization priority combinations, allowing for indirect Pareto surface inference.

The model’s performance was analyzed on conventional prostate volumetric modulated arc therapy (VMAT), pancreas stereotactic body radiation therapy (SBRT), and spine stereotactic radiosurgery (SRS) with epidural space irradiation. For each of these treatment paradigms, the Pareto surfaces of many patients were thoroughly sampled to train and test the model. On all of these cases, our model achieved good performance in terms of speed and accuracy. Overfitting was shown to be minimal in all cases, and dose distribution slices and dose-volume histograms (DVHs) were shown for comparison, confirming the proficiency of our model. This model is relatively fast (0.05-0.20 seconds per plan), and it is capable of sampling the entire Pareto surface much faster than commercial dose optimization and calculation engines.

While these results were generally promising, the model achieved lower error on the prostate VMAT treatment plans compared to the pancreas SBRT and spine SRS treatment plans. This is likely due to the existence of heavier beam streaks in the stereotactic treatment plans which are generated by a sharper control of the delivered dose distribution. However, the Pareto surface errors were similar across all three cases, so these dose distribution errors did not propagate to the Pareto objective space.

The secondary focus of this study is the development and analysis of Pareto surface similarity metrics. The dose prediction model can be used to rapidly estimate many Pareto-optimal plans for quick Pareto surface inference. This could allow for a potentially significant increase in the speed at which Pareto surfaces are inferenced to provide treatment planning assistance and acceleration. However, previous investigations into Pareto surface analysis typically do not compare a ground truth Pareto surface with a Pareto surface prediction. Therefore, there is a need to develop a Pareto surface metric in order to evaluate the ability of the model to generate accurate Pareto surfaces in addition to accurate dose distributions.

To address these needs, we developed four Pareto surface similarity metrics, emphasizing the ability to represent distances between the interpolations rather than the sampled points. The most straightforward metric is the root-mean-square error (RMSE) evaluated between matched, sampled points on the Pareto surfaces, augmented by intra-simplex upsampling of the barycentric dimensions of each simplex. The second metric is the Hausdorff distance, which evaluates the maximum closest distance between the sets of sampled points. The third metric is the average projected distance (APD), which evaluates the displacements between the sampled points and evaluates their projections along the mean displacement. The fourth metric is the average nearest-point distance (ANPD), which numerically integrates point-to-simplex distances over the upsampled simplices of the Pareto surfaces. These metrics are compared by their convergence rates as a function of intra-simplex upsampling, the calculation times required to achieve convergence, and their qualitative meaningfulness in representing the underlying interpolated surfaces. For testing, several simplex pairs were constructed abstractly, and Pareto surfaces were constructed using inverse optimization and our dose prediction model applied to conventional prostate VMAT, pancreas SBRT, and spine SRS with epidural irradiation.

For the abstract simplex pairs, convergence within 1% was typically achieved at approximately 50 and 100 samples per barycentric dimension for the ANPD and the RMSE, respectively. The RMSE and the ANPD required approximately 50 milliseconds and 3 seconds to calculate to these sampling rates, respectively, while the APD and HD required much less than 1 millisecond. Additionally, the APD values closely resembled the ANPD limits, while the RMSE limits and HD tended to be more different. The ANPD is likely more meaningful than the RMSE and APD, as the ANPD’s point-to-simplex distance functions more closely represent the dissimilarity between the underlying interpolated surfaces rather than the sampling points on the surfaces. However, in situations requiring high-speed evaluations, the APD may be more desirable due to its speed, lack of subjective specification of intra-simplex upsampling rates, and similarity to the ANPD limits.

The tertiary focus of this study is the analysis of the feasibility of epidural space irradiation in spine SRS. The epidural space is a frequent site of cancer recurrence after spine SRS. This may be due to microscopic disease in the epidural space which is under-dosed to obey strict spinal cord dose constraints. We hypothesized that the epidural space could be purposefully irradiated to prescription dose levels, potentially reducing the risk of recurrence in the epidural space without increasing toxicity. To address this, we sought to analyze the feasibility of irradiating the epidural space in spine SRS. Analyzing the data associated with this study is synergistic to our MCO acceleration study, since the range of trade-offs between epidural space irradiation and spinal cord sparing represents an MCO problem which our dose prediction model may quickly solve.

Spine SRS clinical treatment plans with associated spinal PTV (PTVspine) and spinal cord contours, and prior delivered dose distributions were identified retrospectively. An epidural space PTV (PTVepidural) was contoured to avoid the spinal cord and focus on regions near the PTVspine. Clinical plan constraints included PTVspine constraints (D95% = 1800 cGy, D5% < 1950 cGy) and spinal cord constraints (Dmax < 1300 cGy, D10% < 1000 cGy). Prior clinical plan doses were mapped onto the new PTVepidural contour for analysis. Plans were copied and revised to additionally target the PTVepidural, optimizing PTVepidural D95% after meeting clinical plan constraints. Tumor control probabilities (TCPs) were estimated for the PTVepidural using a radiobiological linear-quadratic model of cell survival for both clinical and revised plans. Clinical and revised plans were compared according to their PTVepidural DVH distributions, D95% distributions, and TCPs.

Seventeen SSRS plans were identified and included in this study. Revised plan DVHs demonstrated higher doses to the epidural low-dose regions, with D95% improving from 10.96 Gy ± 1.76 Gy to 16.84 Gy ± 0.87 Gy (p < 10-5). Our TCP modeling set the clinical plan TCP average to 85%, while revised plan TCPs were all greater than 99.99%. Therefore, irradiating the epidural space in spine SRS is likely feasible, and purposefully targeting the epidural space in SSRS should increase control in the epidural space without significantly increasing the risk of spinal cord toxicity.

Artificial intelligence (AI) has been rapidly developing in various fields, featuring automation in complex tasks with superior efficiency. This feature meets the urgent need for the automation of resource-intensive tasks in clinics. In radiation oncology, AI has been investigated for almost every process in patient management and treatment. Among these, radiotherapy treatment planning is one of the most time-consuming and labor-intensive processes. This dissertation work focuses on AI-based planning agents for intensity-modulated radiation therapy (IMRT) for various treatment sites. Fluence map prediction for prostate simultaneous integrated boost (SIB) or Stereotactic Body Radiotherapy (SBRT) cases was selected for a feasibility study. Prostate cases have one of the most consistent anatomic geometries and dosimetric constraints among all treatment sites. The developed prostate AI planning agent employed a customized convolutional neuro network (CNN), Dense-Res Hybrid Network (DRHN). DRHN was trained to predict optimal fluence maps from patient anatomic information. The proposed method avoids the time-consuming inverse planning process and thus could make fluence map predictions in seconds and generate IMRT plans in a few minutes. The resulting AI plan quality met institutional clinical guidelines. This preliminary study demonstrated the feasibility of the proposed AI strategy in automatic treatment planning and provided a solid foundation for the following studies. As a step forward, a more sophisticated AI agent for oropharyngeal cases was developed based on the prostate AI agent. This AI agent had the following two upgrades to adapt to the much more complex geometry in head-and-neck (H&N) treatment site: 1) conditional generative adversarial networks (cGAN) training architecture; 2) the generator, PyraNet, was a customized CNN network with more complicated network structure design in the shape of pyramids. This H&N AI agent demonstrated encouraging plan quality, especially that organs-at-risk (OAR) dosimetric outcomes achieved expectations. A graphical user interface (GUI) was developed and commissioned to make this AI tool available for clinical implementation. In summary, a DL-based fluence map prediction was developed for prostate and H&N cases. The H&N AI agent was implemented for clinical use, and more related research and applications are around the corner.

With the development of medical linear accelerator technologies, the precision and complexity of external beam radiation therapy have increased tremendously over the years. The goal of radiation therapy has always been to push the limit to irradiate the target volume while preserving normal tissues. To achieve this goal, treatment planning for radiation therapy has become a labor-intensive and time-consuming task, which requires a high level of experience and knowledge from the planner. Therefore, automated treatment planning, or auto-planning, is of particular interest in radiation therapy research. The advantages of auto-planning are reduced planning time and increased plan quality consistency.Since the treatment planning workflow has multiple steps, auto-planning includes the automation of different planning procedures, such as contouring, beam placement, and inverse optimization, which can be achieved in different approaches. The main approaches are knowledge-based planning, automated rule implementation and reasoning, and multicriteria optimization. We can generally consider such novel auto-planning applications as artificial intelligence (AI). This study primarily focuses on treatment plan generation using knowledge-based planning and machine learning techniques. The study includes two main projects: automated beam setting for whole breast radiation therapy (WBRT) and fluence map prediction for intensity modulated radiation therapy (IMRT). In WBRT planning, tangential beams are used to irradiate the entire breast volume and avoid the organs-at-risk (OARs) (i.e., the lungs and the heart) as much as possible. The placement of the beams is vital in determining the planning target volume (PTV) coverage and normal tissue sparing. Furthermore, planners need to take multiple clinical considerations into account, e.g., avoiding the contralateral breast and the heart, and use a variety of techniques to meet the demands. Therefore, we developed an automated beam setting program which takes simple user settings and optimizes target coverage and OAR sparing. The program can be launched from the Eclipse Treatment Planning System (TPS) as a binary plug-in script, which generates a graphical user interface to accept user inputs. Several beam geometries are supported: tangential beams only (supine), tangential plus supraclavicular (SCV) beams (supine), and prone beams. For all geometries, the program calculates the optimal gantry angles, collimator angles, isocenter location, jaw sizes, and MLC shapes. The borders of the SCV beams are also matched to the tangential beams by using couch kicks on the main tangential fields. For the supine geometries, a coefficient was learned from existing clinical plans to balance between the PTV and lung coverages. The program searches from an initial setting based on breast wires and finds the optimal setting. For the prone geometry, the planner can set a margin to customize the coverage near the PTV-lung interface. The program has been implemented together with a WBRT fluence prediction program, which creates electronic compensation (ECOMP) plans from the given beam settings. This automated workflow can significantly reduce the workload of the forward planned ECOMP plans. The results showed that the AI plans achieved similar or better plan quality compared to the manual plans. In IMRT planning, inverse optimization is the standard practice to create treatment plans. Dose-volume histogram (DVH) constraints and priorities are set by the planner to start the optimization and often continuously tuned throughout the planning process until the optimal dose distribution is achieved. The actual parameters to be optimized are fluence map intensities of the IMRT beams. Numerous efforts have been devoted in KBP to predict either the DVH or the dose of the optimal plan. The rationale is that, given the patient anatomy and the physician’s prescription, the DVH or dose in the final plan can be predicted based on similar previous plans. The predicted DVH or dose can then be used as a reference to either evaluate the plan quality or generate new plans by converting them into inverse optimization objectives, which is a process also known as dose mimicking. However, most dose mimicking techniques are still in the development stage and not yet commercially available. We explored the feasibility to directly predict optimal fluence maps and generate IMRT plans without inverse optimization. In order to achieve fluence map prediction, we first investigated the correlation between patient anatomy and fluence maps. A database of patient anatomy and fluence maps was built with pancreas SBRT cases. Treatment planning was done on 2D axial slices with in-house dose calculation and fluence optimization algorithms. For a new slice, an atlas matching method was developed to search for the most anatomically similar slice in the database and initialize the optimization with the existing fluence. The atlas-guided fluence optimization reduced the optimization cost and offered a small dosimetric improvement compared to uniform initialization. With more training data, deep learning methods were experimented to predict fluence maps from patient anatomy. A deep learning framework consisting of two convolutional neural networks (CNN) was developed. As each plan has several beams, all beam doses must add up to the optimal plan’s total dose, while each beam dose is deposited only by said beam’s fluence map. Therefore, the BD-CNN predicts the individual beam doses (BD) for an IMRT plan, which tries to minimize the prediction error for both the beam doses and the total dose. Once the beam doses are available, each fluence map (FM) is generated separately by the FM-CNN. As the fluence maps exist in the beam’s eye view (BEV), a projection of the 3D beam dose onto the 2D BEV is necessary. The resulting dose map is used as the input to the FM-CNN, which predicts the fluence map as the output. The predicted fluence maps are imported into the TPS for leaf sequencing and dose calculation, generating a deliverable plan. These projects are retrospective studies using anonymized patient data for training and testing. The development of the deep learning framework was split into several stages: the initial test of the feasibility was conducted for pancreas stereotactic body radiation therapy (SBRT) with a single PTV, unified dose constraints, and a fixed 9-beam geometry; the networks were then modified to allow variable dose inputs and multiple PTVs for pancreas SBRT with simultaneous integrated boost (SIB); a transfer learning technique was applied to the training of the framework for adrenal SBRT plans with different beam settings and dose constraints, using the pancreas model as the base model. The framework has evolved to be more robust and support different sites and planning styles over time. The AI plans with predicted fluence maps achieved similar plan quality as manual plans for most cases. For some cases with particularly challenging patient anatomies, the AI plans can struggle to reach the high standard of the expert plans. Fluence map prediction is a viable way to directly generate IMRT plans without inverse optimization. This application may be especially useful for adaptive treatment planning.

Purpose: To develop a heuristic greedy algorithm to generate automatic multi-leaf collimator (MLC) sequencing for spatial fractioned radiation therapy (SFRT) using 3D dynamic conformal arc (DCA).Methods and materials: One late stage lung cancer patient with simulated sphere target grid was included in this study. N_t spheres were equally spaced within the gross target volume (GTV). The sphere targets are 1.5 cm in diameter, 4.3 cm spacing for 6,9,10, and 12 targets scenarios, and 2.8cm spacing for one special 10 targets scenario. Optimization was designed to complete within one coplanar arc from 180° to 0° in a clockwise direction with 2° as the angle interval. The problem is formalized as finding optimal MLC sequencing to cover N_t targets with K control points (CPs) for each arc. The state of each target’s MLC opening at each CP is binary. The original NP-hard problem can be approximated to a feasible subproblem by the greedy approximation on each control point and using the heuristic approach for the initial point. The algorithm focuses on the normalized relative dose relationship as the object function during the optimization. The dose matrix for each step was rasterized and grouped based on Monte Carlo simulation as the pre-calculation process. The physical speed limitation of the MLC motion was considered in the optimization to achieve a realistic and deliverable final MLC sequencing solution. Four grid arrays (6, 9, 10, and 12 targets respectively) were tested for plan quality. The arc collimator angle was planned with both 0 and 30 degrees for comparison. Prescription was set to 20 Gy to one fraction. The delivered dose will be normalized to equalize the minimum target dose to the prescription dose. Key dosimetric endpoints including target mean dose, D5, and D95, were reported. Results: The complexity of this algorithm has been reduced by a factor of \frac{2^K}{2\left(K-1\right)}. The D95 deviations of all targets as the main focus object were within 2.88% in four grid arrays with 0°/30° collimator rotation angles, 4.3 cm spacing for 6,9,10 and 12 targets scenarios, and 2.8 cm spacing for one special 10 targets scenario. For all scenarios with 4.3 cm spacing, the mean valley-to-peak ratios were under 0.45 and were within the constraint that the dose of the other part of the tumor is no more than 45% of the max normalized D95 delivered target dose during the algorithm optimization. Conclusion: This algorithm is a feasible and practical method with high efficiency while delivering the prescription dose to small target volume for late stage cancer palliative management. The proposed solution provides decent coverage to the tumor volume as well as the valley-to-peak ratio. It provides a competitive alternative solution to the standard alloy grid delivery technique.

Purpose: To investigate the effect of incorporating a beam spreading parameter in a beam angle optimization algorithm and to evaluate its efficacy for creating coplanar IMRT lung plans in conjunction with machine learning generated dose objectives.

Methods: Fifteen anonymized patient cases were each re-planned with ten values over the range of the beam spreading parameter, k, and analyzed with a Wilcoxon signed-rank test to determine whether any particular value resulted in significant improvement over the initially treated plan created by a trained dosimetrist. Dose constraints were generated by a machine learning algorithm and kept constant for each case across all k values. Parameters investigated for potential improvement included mean lung dose, V20 lung, V40 heart, 80% conformity index, and 90% conformity index.

Results: With a confidence level of 5%, treatment plans created with this method resulted in significantly better conformity indices. Dose coverage to the PTV was improved by an average of 12% over the initial plans. At the same time, these treatment plans showed no significant difference in mean lung dose, V20 lung, or V40 heart when compared to the initial plans; however, it should be noted that these results could be influenced by the small sample size of patient cases.

Conclusions: The beam angle optimization algorithm, with the inclusion of the beam spreading parameter k, increases the dose conformity of the automatically generated treatment plans over that of the initial plans without adversely affecting the dose to organs at risk. This parameter can be varied according to physician preference in order to control the tradeoff between dose conformity and OAR sparing without compromising the integrity of the plan.

Purpose: To develop an automated program that can generate the optimal beams for whole breast radiation therapy (WBRT).

Methods and Materials: A total of twenty patients receiving WBRT were included in this study. The computed tomography (CT) simulation images and structures of all 20 patients were used to develop and validate the program. All patients had the breast planning target volume (PTV) contour drawn by physicians and radio-opaque catheters placed on the skin during CT simulation. First, an initial beam was calculated based on the CT images, the radio-opaque catheters, and the breast PTV contour. The beam includes five main parameters: the gantry angles, the isocenter location, the field size, the collimator angles, and the initial multi-leaf collimator (MLC) shape.

To optimize the beam parameters, a geometry-based objective function was constructed to optimize target coverage and organ-at-risk (OAR) sparing. The objective function is the weighted sum of the square of the relative volumes of the PTV outside the field and the ipsilateral lung inside the field. Due to the curvature of the chest wall, a portion of the ipsilateral lung will be included in the irradiated volume. The balance between PTV coverage and OAR sparing is embodied by the relative weight of the lung volume in the objective function, which was trained and validated from the clinical plans of the twenty patients. Two different optimization schemes were developed to minimize the objective function: the exhaustive search and the local search. The search was conducted in a 2-dimensional grid with the gantry angle (1° increments) and the isocenter location (1 mm increments) as two axes and the initial beam as the origin point. For the exhaustive search, the ranges of the gantry angle and the isocenter location are ±12° and ±21 mm. The local search does not require a search range. The beam with the minimal objective function value in the grid is considered optimal. The optimal beam was transferred to an in-house automatic fluence optimization program developed specifically for WBRT. The automatic plans were compared with the manually generated clinical plans for target coverage, dose conformity and homogeneity, and OAR dose.

Results: The calculation time of the beam optimization was under one minute for all cases. The local search (~15 s) took less time than the exhaustive search (~45 s), and the two methods produced the same result for the same patient. The automatic plans have overall comparable plan quality to the clinical plans, which usually take 1 to 4 hours to make. Generally, the PTV coverage is improved while the dose to the ipsilateral lung and the heart is similar. The breast PTV Eval V95% of all cases are above 95%, and the mean V95% (97.7%) is increased compared with the clinical plans (96.8%). The ipsilateral lung V16Gy is reduced for 14 out of 20 cases, and the mean V16Gy is decreased in the automatic plans (12.6% vs. 13.6%). The average heart mean dose is slightly increased in the automatic plans (2.06% vs. 1.99%).

Conclusion: Optimal beams for WBRT can be automatically generated in one minute given the patient’s simulation CT images and structures. The automated beam setup program offers a valuable tool for WBRT planning, as it provides clinically relevant solutions based on previous clinical practice as well as patient specific anatomy.

Knowledge-based radiation treatment is an emerging concept in radiotherapy. It

mainly refers to the technique that can guide or automate treatment planning in

clinic by learning from prior knowledge. Dierent models are developed to realize

it, one of which is proposed by Yuan et al. at Duke for lung IMRT planning. This

model can automatically determine both beam conguration and optimization ob-

jectives with non-coplanar beams based on patient-specic anatomical information.

Although plans automatically generated by this model demonstrate equivalent or

better dosimetric quality compared to clinical approved plans, its validity and gener-

ality are limited due to the empirical assignment to a coecient called angle spread

constraint dened in the beam eciency index used for beam ranking. To eliminate

these limitations, a systematic study on this coecient is needed to acquire evidences

for its optimal value.

To achieve this purpose, eleven lung cancer patients with complex tumor shape

with non-coplanar beams adopted in clinical approved plans were retrospectively

studied in the frame of the automatic lung IMRT treatment algorithm. The primary

and boost plans used in three patients were treated as dierent cases due to the

dierent target size and shape. A total of 14 lung cases, thus, were re-planned using

the knowledge-based automatic lung IMRT planning algorithm by varying angle

spread constraint from 0 to 1 with increment of 0.2. A modied beam angle eciency

index used for navigate the beam selection was adopted. Great eorts were made to assure the quality of plans associated to every angle spread constraint as good

as possible. Important dosimetric parameters for PTV and OARs, quantitatively

re

ecting the plan quality, were extracted from the DVHs and analyzed as a function

of angle spread constraint for each case. Comparisons of these parameters between

clinical plans and model-based plans were evaluated by two-sampled Students t-tests,

and regression analysis on a composite index built on the percentage errors between

dosimetric parameters in the model-based plans and those in the clinical plans as a

function of angle spread constraint was performed.

Results show that model-based plans generally have equivalent or better quality

than clinical approved plans, qualitatively and quantitatively. All dosimetric param-

eters except those for lungs in the automatically generated plans are statistically

better or comparable to those in the clinical plans. On average, more than 15% re-

duction on conformity index and homogeneity index for PTV and V40, V60 for heart

while an 8% and 3% increase on V5, V20 for lungs, respectively, are observed. The

intra-plan comparison among model-based plans demonstrates that plan quality does

not change much with angle spread constraint larger than 0.4. Further examination

on the variation curve of the composite index as a function of angle spread constraint

shows that 0.6 is the optimal value that can result in statistically the best achievable

plans.

Purpose: To develop and evaluate an automatic IMRT planning technique for prostate cancer utilizing prior expert plan's dose distribution as guidance.

Methods and Materials: In this study, the anatomical information of prostate cancer cases was parameterized and quantified into two measures: the percent distance-to-prostate (PDP) and the concaveness angle. Based on these two quantities, a plan atlas composed of 5 expert prostate IMRT plans was built out of a 70-case pool at our institution using k-medoids clustering analysis.

Extra 20 cases were used as query cases to evaluate the dose-guided automatic planning (DAP) scheme. Each query case was matched to an atlas case based on PTV-OAR anatomical features followed by deformable registration to enhance fine local matching. Using the deformation field, the expert dose in the matched atlas case was warped onto the query case, creating the goal dose conformal to the query case's anatomy. Dose volume histograms (DVHs) objectives were sampled from the goal dose to guide automatic IMRT treatment planning. Dosimetric comparison between DAP plans and clinical plans were performed.

Results: Generating goal dose is highly efficient by using MIM^TM workflows. The deformable registration provides high-quality goal dose tailored to query case's anatomy in terms of the dose falloff at the PTV-OAR boundary and the overall conformity. Automatic planning in Eclipse^TM takes ~2.5 min (~70 iterations) without human intervention. Compared to clinical plans, DAP plans improved the conformity index from 0.85±0.04 to 0.88±0.02 (p=0.0045), the bladder-gEUD from 40.7±3.2 Gy to 40.0±3.1 Gy (p=0.0003), and rectum-gEUD from 40.4±2.0 Gy to 39.9±2.1 Gy (p=0.0167). Other dosimetric parameter is similar (p>0.05): homogeneity indices are 7.4±0.9% and 7.1±1.5%, for DAP plans and clinical plans, respectively.

Conclusions: Dose-guided automatic treatment planning is feasible and efficient. Atlas-based patient-specific dose objectives can effectively guide the optimizer to achieve similar or better plan quality compared to clinical plans.

Radiation has been utilized in medical domain for multiple purposes. Treating cancer using radiation has increasing popularity during the last century. Radiation beam is directed to the tumor cells while the surrounding healthy tissue is attempted to be avoided. Radiation therapy treatment planning serves the goal of delivering high concentrated radiation to the treatment volume while minimizing the normal tissue as much as possible. With the advent of more sophisticated delivery technology, treatment planning time increases over time. In addition, the treatment plan quality relies on the experience of the planner. Several computer assistance techniques emerged to help the treatment planning process, among which knowledge-based planning (KBP) has been successful in inverse planning IMRT. KBP falls under the umbrella of Knowledge Discovery in Databases (KDD) which originated in industry. The philosophy is to extract useful knowledge from previous application/data/observations to make predictions in the future practice. KBP reduces the iterative trial-and-error process in manual planning, and more importantly guarantees consistent plan quality. Despite the great potential of treatment planning KDD (TPKDD), three major challenges remain before TPKDD can be widely implemented in the clinical environment: 1. a good knowledge model asks for sufficient amount of training data to extract useful knowledge and is therefore less efficient; 2. a knowledge model is usually only applicable for the specific treatment site and treatment technique and is therefore less generalizable; 3. a knowledge model needs meticulous inspection before implementing in the clinic to verify the robustness.

This study aims at filling in the niche in TPKDD and improves current TPKDD workflow by tackling the aforementioned challenges. This study is divided into three parts. The first part of the study aims to improve the modeling efficiency by introducing an atlas-based treatment planning guidance. In the second part of the study, an automated treatment planning technique for whole breast radiation therapy (WBRT) is proposed to provide a solution for the area where TPKDD has not yet set foot on. In the third part of the study, several topics related to the knowledge model quality are addressed, including improving the model training workflow, identifying geometric novelty and dosimetric outlier case, building a global model and facilitating incremental learning.

I. Improvement of the modeling efficiency. First, a prostate cancer patient anatomy atlas was established to generate 3D dose distribution guidance for the new patient. The anatomy pattern of the prostate cancer patient was parameterized with two descriptors. Each training case was represented in 2D feature space. All training cases were clustered using the k-medoids algorithm. The optimal number of clusters was determined by the largest average silhouette width. For the new case, the most similar case in the atlas was identified and used to generate dose guidance. The anatomy of the atlas case and the query case was registered and the deformation field was applied to the 3D radiation dose of the atlas case. The deformed dose served as the goal dose for the query case. Dose volume objectives were then extracted from the goal dose to guide the inverse IMRT planning. Results showed that the plans generated with atlas guidance had similar dosimetric quality as compared to the clinical manual plans. The monitor units (MU) of the auto plan were also comparable with the clinical plan. Atlas-guided radiation therapy has proven to be effective and efficient in inverse IMRT planning.

II. Improvement of model generalization. An automatic WBRT treatment planning workflow was developed. First of all, an energy selection tool was developed based on previous single energy and dual energy WBRT plans. The DRR intensity histograms of training cases were collected and the principal component analysis (PCA) was performed to reduce the dimension of the histogram. First two components were used to represent each case and the classification was performed in the 2D space. This tool helps new patient to select appropriate energy based on the anatomy information. Secondly, an anatomy feature based random forest (RF) model was proposed to predict the fluence map for the patient. The model took the input of multiple anatomical features and output the fluence intensity of each pixel within the fluence map. Finally, a physics rule based method was proposed to further fine tune the fluence map to achieve optimal dose distribution within the irradiated volume. Extra validation cases were tested on the proposed workflow. Results showed similar dosimetric quality between auto plan and clinical manual plan. The treatment planning time was reduced from between 1-4 hours for the manual planning to within 1 minute for the auto planning. The proposed automatic WBRT planning technique has proven to be efficient.

III. Rapid learning of radiation therapy KBP. Several topics were analyzed in this part of the study. First of all, a systematic workflow was established to improve the KBP model quality. The workflow started with identifying geometric novelty case using the statistical metric “leverage”, followed by removing the novelty case. Then the dosimetric outlier was identified using studentized residual and then cleaned. The cleaned model was compared with the uncleaned model using the extra validation cases. This study used pelvic cases as an example. Results showed that the existence of novelty and outlier cases did degrade the model quality. The proposed statistical tools can effectively identify novelty and outlier cases. The workflow is able to improve the quality of the knowledge-based model.

Secondly, a clustering-based method was proposed to identify multiple geometric novelty cases and dosimetric outlier cases at the same time. One class support vector machine (OCSVM) was applied to the feature vectors of all training cases to generate one class of inliers while cases falling out of the frontier belonged to the novelty case group. Once the novelty cases were identified and cleaned, the robust regression followed by outlier identification (ROUT) was applied to all remaining cases to identify dosimetric outliers. A cleaned model was trained with the novelty and outlier free case pool and was tested using 10 fold cross validation. Initial training pool included intentionally added outlier cases to evaluate the efficacy of the proposed method. The model prediction on the inlier cases was compared with that of novelty and outlier cases. Results showed that the method can successfully identify geometric novelty and dosimetric outliers. The model prediction accuracy between the inliers and novelty/outliers was significantly different, indicating different dosimetric behavior between two groups. The proposed method proved to be effective in identifying multiple geometric novelty and dosimetric outliers.

Thirdly, a global model using the model tree and the clustering-based model was proposed to include cases with different clinical conditions and indications. The model tree is a combination of decision tree and linear regression, where all cases are branched into leaves and regression is performed within each leaf. A clustering-based model used k-means algorithm to segment all cases into more aggregated groups, and then the regression was performed within each small group. The overall philosophy of both the model tree and the clustering-based method is that cases with similar features have similar geometry-dosimetry relation. Training cases within small feature range gives better model accuracy. The proposed method proved to be effective in improving the model accuracy over the model trained with all cases without segmenting the cases.

At last, the incremental learning was analyzed in radiation therapy treatment planning model. This study tries to answer the question when model re-training should be invoked. In the clinical environment, it is often unnecessary to re-train the model whenever there is a new case. The scenario of incrementally adapting the model was simulated using the pelvic cases with different number of training cases and new incoming cases. The result showed that re-training was often necessary for small training dataset and as the number of cases increased, re-training became less frequent.

In summary, this study addressed three major challenges in TPKDD. In the first part, an atlas-guided treatment planning technique was proposed to improve the modeling efficiency. In the second part, an automatic whole breast radiation therapy treatment planning technique was proposed to tackle the issue where TPKDD has not yet resolved. In the final part, outlier analysis, global model training and incremental learning was further analyzed to facilitate rapid learning, which lay the foundation of future clinical implementation of radiation therapy knowledge models.

Purpose: To build a model that will predict the survival time for patients that were treated with stereotactic radiosurgery for brain metastases using support vector machine (SVM) regression.

Methods and Materials: This study utilized data from 481 patients, which were equally divided into training and validation datasets randomly. The SVM model used a Gaussian RBF function, along with various parameters, such as the size of the epsilon insensitive region and the cost parameter (C) that are used to control the amount of error tolerated by the model. The predictor variables for the SVM model consisted of the actual survival time of the patient, the number of brain metastases, the graded prognostic assessment (GPA) and Karnofsky Performance Scale (KPS) scores, prescription dose, and the largest planning target volume (PTV). The response of the model is the survival time of the patient. The resulting survival time predictions were analyzed against the actual survival times by single parameter classification and two-parameter classification. The predicted mean survival times within each classification were compared with the actual values to obtain the confidence interval associated with the model’s predictions. In addition to visualizing the data on plots using the means and error bars, the correlation coefficients between the actual and predicted means of the survival times were calculated during each step of the classification.

Results: The number of metastases and KPS scores, were consistently shown to be the strongest predictors in the single parameter classification, and were subsequently used as first classifiers in the two-parameter classification. When the survival times were analyzed with the number of metastases as the first classifier, the best correlation was obtained for patients with 3 metastases, while patients with 4 or 5 metastases had significantly worse results. When the KPS score was used as the first classifier, patients with a KPS score of 60 and 90/100 had similar strong correlation results. These mixed results are likely due to the limited data available for patients with more than 3 metastases or KPS scores of 60 or less.

Conclusions: The number of metastases and the KPS score both showed to be strong predictors of patient survival time. The model was less accurate for patients with more metastases and certain KPS scores due to the lack of training data.

The aim of the study is to develop a geometrically adaptive and statistically robust plan quality inference method. A knowledge-based plan quality inference method is proposed in this study. It references to similar plans in the history database for patient-specific plan quality evaluation. Similar plans are retrieved using a novel plan similarity metric, and dosimetric statistical inferences are obtained from the selected similar plans. Two plan quality metrics—dosimetric result probability (DRP) and dose deviation index (DDI)—are proposed to quantify plan quality amongst prior similar plans. 927 clinical approved head-and-neck treatment plans with two planning targets were exported and used as the historical database. Eight organs-at-risk (OARs), including brainstem, spinal cord, larynx, mandible, pharynx, oral cavity, left parotid and right parotid were analyzed in this study. Statistical analysis is performed to validate the similarity of the selected reference plans. 12 sub-optimal plans identified by DRP were re-planned to validate the capability of the proposed methods in identifying inferior plans, To demonstrate the potential of our proposed method as a plan quality data analytics tool, a population-wise analysis was conducted on all retrieved plans sorted every two years. A ready-to-use stand-along application was also developed to streamline the evaluation process.

After replanning, left and right parotid median dose are reduced by 31.7\% and 18.2\% respectively. 83\% of these cases would not be identified as suboptimal without the proposed similarity plan selection. The population plan quality analysis reveals that the average parotid sparing has increased by 21.7\% from 2005 to 2018. Notably, the increasing dose sparing over time in retrospective plan quality analysis is strongly correlated with the increasing dose prescription ratios to the two planning targets, revealing the collective trend in planning conventions.

The proposed similar plan retrieval and analysis methodology has been proven to be predictive of the current plan quality. Therefore, the proposed workflow can potentially be applied in the clinics as a real-time plan quality assurance tool. The proposed metrics can also serve the purpose of plan quality analytics in finding connections and historical trends in the clinical treatment planning workflow.

Abstract

Intensity Modulated Radiation Therapy (IMRT) and Volumetric Modulated Arc Therapy (VMAT) have become effective tools for treating cancer with radiation. Designing a high quality IMRT/VMAT treatment plan is time consuming. Different kinds of knowledge-based methods are being developed to reduce planning time and improve the plan quality by extracting knowledge from previous expert plans to form knowledge models and applying such models to the new patient cases. Currently, these methods are mostly limited to a particular cancer type and therefore various diseases types require training of multiple knowledge models with a large number of cases.

To investigate the feasibility of knowledge modeling of IMRT/VMAT treatment planning for multiple cancer types, a progressive study is conducted with a treatment planning knowledge model that quantifies correlations between patient pelvic anatomical features and the OAR sparing features. Low risk prostate plans with relatively simpler PTV-OAR geometry, which is the most common geometry type in previous knowledge based studies, are used to train the model as the starting point of the progressive modeling process. Cases with more complex PTV-OAR anatomies (prostate cancer cases with lymph node irradiation, and anal rectal cancer cases) are added to the training dataset one by one until the model prediction accuracies reach plateau. The DVHs predicted by the knowledge model for bladder, femoral heads and rectum are validated by cases from all three types of cases. Dosimetric parameters are extracted from the predicted DVHs and the corresponding actual plan values measure the prediction accuracy of this multi-disease type model. Further, its accuracy was also compared with the models trained by single disease type cases (including low risk prostate cancer, or type 1, high risk prostate cancer with lymph nodes, or type 2 and anal rectal cancer, or type 3).

Prediction accuracy reaches plateau when 6 high risk prostate cancer with lymph nodes irradiation cases and 8 anal rectal cancer cases were added to the training dataset. The determination coefficients R2 for the OARs are: Bladder: 0.90, rectum: 0.64 and femoral heads: 0.82. The prediction accuracies by the multi-disease type model and single-disease type models have no significant differences by F-test (p-value: bladder: 0.58, femoral head: 0.44, rectum: 0.97).

Conclusion:

Progressive knowledge modeling of OAR sparing for multiple cancer types in in the pelvic region is feasible and has comparable accuracy to single-disease type modeling.

Purpose: The purpose of this study is to utilize F-18 FDG PET/CT scans to determine an indicator for the response of esophageal cancer patients during radiation therapy. There is a need for such an indicator since local failures are quite common in esophageal cancer patients despite modern treatment techniques. If an indicator is found, a patient's treatment strategy may be altered to possibly improve the outcome. This is investigated with various standard uptake volume (SUV) metrics along with image texture features. The metrics and features showing the most promise and indicating response are used in logistic regression analysis to find an equation for the prediction of response.

Materials and Methods: 28 patients underwent F-18 FDG PET/CT scans prior to the start of radiation therapy (RT). A second PET/CT scan was administered following the delivery of ~32 Gray (Gy) of dose. A physician contoured gross tumor volume (GTV) was used to delineate a PET based GTV (GTV-pre-PET) based on a threshold of >40% and >20% of the maximum SUV value in the GTV. Deformable registration was used in VelocityAI software to register the pre-treatment and intra-treatment CT scans so that the GTV-pre-PET contours could be transferred from the pre to intra scans (GTV-intra-PET). The fractional decrease in the maximum, mean, volume to the highest intensity 10%-90%, and combination SUV metrics of the significant previous SUV metrics were compared to post-treatment pathologic response for an indication of response. Next for the >40% threshold, texture features based on a neighborhood gray-tone dimension matrix (NGTDM) were analyzed. The fractional decrease in coarseness, contrast, busyness, complexity, and texture strength were compared to the pathologic response of the patients. From these previous two types of analysis, SUV and texture features, the two most significant results were used in logistic regression analysis to find an equation to predict the probability of a non-responder. These probability values were then used to compare against the pathological response to test for indication of response.

Results:

20 of the 28 patients underwent post treatment surgery and their pathologic response was determined. 9 of the patients were classified as being responders (treatment effect grade ≤ 1) while 11 of the patients were classified as being non-responders (treatment effect grade > 1). The fractional difference in the different SUV metrics has shown that the most commonly used maximum SUV and mean SUV were not significant in determining response to the treatment. Other SUV metrics however did show promise as being indicators. For the >40% threshold SUV to the highest 10%, 20%, and 30% (SUV10%, SUV20%, SUV30%) were found to significantly distinguish between responders and non-responders (p=0.004) and had an area under the Receiver Operating Characteristic curve (AUC) of 0.7778. Combining these significant metrics (SUV10% with SUV20% and SUV 20% with SUV30%) also was able to distinguish response (p=0.033, AUC=0.7879). Cross validation of these results shown that these metrics could be used to find the response on previously unseen data. The three individual SUV terms distinguished responders from non-responders with a sensitivity of 0.7143 and a specificity of 0.6400 from the cross validation. Cross validation yielded a sensitivity of 0.8333 and a specificity of 0.7727 for the combination of SUV10% and SUV20% and a sensitivity of 0.8333 and specificity of 0.7273 for the combination of SUV20% and SUV30%. For the >20% threshold two SUV metrics were found to be significant. These were the SUV to the highest 10% and 20% (p=0.0048). The AUC for the 10% metrics was 0.7677 and for the 20% metric it was 0.7374. Cross validation of these two metrics shown that the 10% metric was the better indicator with being able to distinguish response in unseen data with a sensitivity of 0.7778 and a specificity of 0.7727.

The only texture feature that was able to determine response was complexity (p-0.04, AUC=0.7778). This metric was no more significant than the three individual SUV metrics but less significant than both of the combination metrics. As with the SUV metrics, cross validation was able to show the robustness of these results. Cross validation yielded a result that could accurately distinguish a response with a sensitivity of 0.8333 and a specificity of 0.7273. Logistic regression fit with features of the two most significant results (complexity and combination of SUV10% with SUV20%) yielded the most significant result (p=0.004. AUC=0.8889). Cross validation of this model resulted in a sensitivity of 0.7982 and a specificity 0.7940. This shows that the model would accurately predict the response to unseen data.

Conclusions:

This study revealed that previously used SUV metrics, maximum and mean SUV, may have to be rethought about being used to determine a response in esophageal cancer patients. The most promising SUV metric was a combination of the SUV10% and SUV20% metric for a GTV created from a threshold of >40% of the maximum SUV value, while the most significant texture feature was complexity. The overall best indicator was the logistic regression fit of the significant metrics of complexity and combination of SUV10% with SUV20%. This was able to distinguish responders from non-responders with a threshold of 0.3186 (sensitivity=0.9091, specificity=0.7778).

The online adaptive radiation therapy for prostate cancer based on re-optimization has been shown to provide better daily target coverage through the treatment course, especially in treatment sessions with large anatomical deformation. However, the clinical implementation of such technique is still limited primarily due to two major challenges: the low efficiency of re-optimization and the lack of online quality assurance technique to verify delivery accuracy. This project aims at developing new techniques and understandings to address these two challenges.

The study was based on retrospective study on patient data following IRB-approved protocol, including both planning Computer Tomography (CT) and daily Cone-Beam Computer Tomography (CBCT) images. The project is divided in to three parts. The first two parts address primarily the efficiency challenge; and the third part of this project aims at validating the deliverability of the online re-optimized plans and developing an online delivery monitoring system.

I. Overall implementation scheme. In this part, an evidence-based scheme, named Adaptive Image-Guided Radiation Therapy (AIGRT), was developed to integrate the re-optimization technique with the current IGRT technique. The AIGRT process first searches for a best plan for the daily target from a plan pool, which consists the original CT plan and all previous re-optimized plans. If successful, the selected plan is used for the daily treatment with translational shifts. Otherwise, the AIGRT invokes re-optimization process of the CT plan for the anatomy-of-the-day, which is added to the plan pool afterwards as a candidate plan for future fractions. The AIGRT scheme is evaluated by comparisons with daily re-optimization and online repositioning techniques based on daily target coverage, Organ-at-Risk (OAR) sparing and implementation efficiency. Simulated treatment courses for 18 patients with re-optimization alone, re-positioning alone and AIGRT shows that AIGRT offers reliable daily target coverage that is highly comparable to re-optimization everyday and significantly improves compared to re-positioning. AIGRT is also seen to provide improved organs-at-risk (OARs) sparing compared to re-positioning. Apart from dosimetric benefits, AIGRT in addition offers an efficient scheme to integrate re-optimization to current re-positioning-based IGRT workflow.

II. Strategies for automatic re-optimization. This part aims at improving the efficiency of re-optimization through automation and strategic selections of optimization parameters. It investigates the strategies for performing fast (~2 min) automatic online re-optimization with a clinical treatment planning system; and explores the performance with different input parameters settings: the DVH objective settings, starting stage and iteration number (in the context of real time planning). Simulated treatments of 10 patients were re-optimized daily for the first week of treatment (5 fractions) using 12 different combinations of optimization strategies. Options for objective settings included guideline-based RTOG objectives, patient-specific objectives based on anatomy on the planning CT, and daily-CBCT anatomy-based objectives adapted from planning CT objectives. Options for starting stages involved starting re-optimization with and without the original plan's fluence map. Options for iteration numbers were 50 and 100. The adapted plans were then analysed by statistical modelling, and compared both in terms of dosimetry and delivery efficiency. The results show that all fast online re-optimized plans provide consistent coverage and conformity to the daily target. For OAR sparing however, different planning parameters led to different optimization results. The 3 input parameters, i.e. DVH objectives, starting stages and iteration numbers, contributed to the outcome of optimization nearly independently. Patient-specific objectives generally provided better OAR sparing compared to guideline-based objectives. The benefit in high-dose sparing from incorporating daily anatomy into objective settings was positively correlated with the relative change in OAR volumes from planning CT to daily CBCT. The use of the original plan fluence map as the starting stage reduced OAR dose at the mid-dose region, but increased 17% more monitor units. Only < 2cc differences in OAR V50% / V70Gy / V76Gy were observed between 100 and 50 iterations. Based on these results, it is feasible to perform automatic online re-optimization in ~2 min using a clinical treatment planning system. Selecting optimal sets of input parameters is the key to achieving high quality re-optimized plans, and should be based on the individual patient's daily anatomy, delivery efficiency and time allowed for plan adaptation.

III. Delivery accuracy evaluation and monitoring. This part of the project aims at validating the deliverability of the online re-optimized plans and developing an online delivery monitoring system. This system is based on input from Dynamic Machine Information (DMI), which continuously reports actual multi-leaf collimator (MLC) positions and machine monitor units (MUs) at 50ms intervals. Based on these DMI inputs, the QA system performed three levels of monitoring/verification on the plan delivery process: (1) Following each input, actual and expected fluence maps delivered up to the current MLC position were dynamically updated using corresponding MLC positions in the DMI. The difference between actual and expected fluence maps creates a fluence error map (FEM), which is used to assess the delivery accuracy. (2) At each control point, actual MLC positions were verified against the treatment plan for potential errors in data transfer between the treatment planning system (TPS) and the MLC controller. (3) After treatment, delivered dose was reconstructed in the treatment planning system based on DMI data during delivery, and compared to planned dose. FEMs from 210 prostate IMRT beams were evaluated for error magnitude and patterns. In addition, systematic MLC errors of ±0.5 and ±1 mm for both banks were simulated to understand error patterns in resulted FEMs. Applying clinical IMRT QA standard to the online re-optimized plans suggests the deliverability of online re-optimized plans are similar to regular IMRT plans. Applying the proposed QA system to online re-optimized plans also reveals excellent delivery accuracy: over 99% leaf position differences are < 0.5 mm, and the majority of pixels in FEMs are < 0.5 MU with errors exceeding 0.5 MU primarily located on the edge of the fields. All clinical FEMs observed in this study have positive errors on the left edges, and negative errors on the right. Analysis on a typical FEM reveals positive correlation between the magnitude of fluence errors and the corresponding leaf speed. FEMs of simulated erroneous delivery exhibit distinct patterns for different MLC error magnitudes and directions, indicating the proposed QA system is highly specific in detecting the source of errors. Based on these results, it can be concluded that the proposed online delivery monitoring system is very sensitive to leaf position errors, highly specific of the error types, and therefore meets the purpose for online delivery accuracy verification. Post-treatment dosimetric verification shows minimal difference between planned and actual delivered DVH, further confirming that the online re-optimized plans can be accurately delivered.

In summary, this project addressed two most important challenges for clinical implementation of online ART, efficiency and quality assurance, through innovative system design, technique development and validation with clinical data. The efficiencies of the overall treatment scheme and the re-optimization process have been improved significantly; and the proposed online quality assurance system is found to be effective in catching and differentiating leaf motion errors.

Purpose: Radiation therapy aims to deliver high energy radiation beam to eradicate cancer cells. Due to radiation toxicity to normal tissue, treatment planning process is needed to customize the radiation beam towards patient specific treatment geometry while minimizing radiation dose to the normal tissue. Treatment planning is often, however, a trial-and-error process to generate ultimate optimal dose distribution. Breast cancer radiation therapy is one of the most commonly seen treatment in modern radiation oncology department. Whole breast radiation therapy (WBRT) using electronic compensation is an iterative manual process which is time consuming. Our institution has been using artificial intelligence (AI) based planning tool for whole breast radiation therapy (WBRT) for 3 years. It is unclear how human planner interacts with AI in real clinical setting and whether the human planner can inject additional insight into well-established AI model. Therefore, the first aim of this study to model planners’ interaction with AI using deep neural network (NN). In addition, we proposed a multi-agent reinforcement learning based framework (MultiRL-FE) to self-interact with the treatment planning system with location awareness to improve plan quality via fluence editing.Methods: A total of 1151 patients have been treated since in-house AI-based planning tool was released for clinical use in 2019. All 526 patients treated with single energy beams were included in this study. The AI tool automatically generates fluence maps and creates “AI plan”. Then planner evaluates the plan and attempts manual fluence modification before physician’s approval (“final plan”). The manual-modification-value (MMV) of each beamlet is the difference between fluence maps in AI and “final plan”. The MMV was recorded for each planner. In the first aim, a deep NN using UNet3+ architecture was developed to predict MMV with AI fluence map, corresponding dose map and organ map in the beam’s eye view (BEV). Then the predicted MMV maps were applied on the initial “AI plan”s to generate AI-modified plans (“AI-m plan”). In the second aim, we developed MultiRL-FE to self-interact with a given plan to improve the plan quality. A simplified treatment planning system was built in the Python environment to train the agent. For each pixel in the fluence map, an individual agent was assigned to interact with the environment by editing fluence value and receive rewards based on projected beam ray’s dose profile. Asynchronous advantage actor critic (A3C) algorithm was used as the backbone for reinforcement learning agents’ training. To effectively train the agent, we developed the MultiRL-FE framework by embedding A3C in a fully convolutional neural network. To test the feasibility of the proposed framework, twelve patients from the same cohort were collected(6 for training and testing respectively). ”Final plans” were perturbed with 10% dose variation to evaluate the potential of the framework to improve the plan. The agent was designed to iteratively modify the fluence maps for 10 iterations. The modified fluence intensity was imported into the Eclipse treatment planning system for dose calculation. For both aims, plan quality was evaluated by dosimetric endpoints including breast PTV V95%(%), V105%(%), V110%(%), lung V20Gy(%) and heart V5Gy(%). Results: In the first aim, the “AI-m plans” generated by HAI network showed statistically significant improvement (p<.05) in hotspot control compared with the initial AI-plan, with an average of -25.2cc volume reduction in breast V105% and -0.805% decrease in Dmax. The planning target volume (PTV) coverage were similar to AI-plan and “final plan”. In the second aim of MultiRL-FE testing, the RL modified plans showed a substantial hotspot reduction from the initial plans. The average PTV V105%(%) of testing set was reduced from 77.78(\pm2.78) to 16.97 (\pm9.42), while clinical plans’ was 3.34(\pm2.73). Meanwhile, the modified plans showed improved dose coverage over the clinical plans, with 70.45(\pm3.94) compared to 65.44(\pm5.39) for V95%(%). Conclusions: In the first part of this study, we proposed a HAI model to enhance the clinical AI tool by reducing hotspot volume from a human perspective. By understanding and modeling the human-automation interaction , this study could advance the widespread clinical application of AI tools in radiation oncology departments with improved robustness and acceptability. In the second part, we developed a self-interactive treatment planning agent with multi-agents reinforcement learning. It offers the advantage of fast location-aware dose editing and can serve as an alternative optimization tool for intensity-modulated radiation therapy and electronic tissue compensation-based treatment planning.

Knowledge-based treatment planning models are commonly built from straightforward principles and utilize experience that it is able to learn from previous high-quality plans. This knowledge can be harnessed by having an e-learning system incorporating knowledge-based treatment planning models to serve as informative, efficient bases to train individuals to develop IMRT plans for a particular site while building confidence in utilizing these models in a clinical setting.

A previously developed beam angle selection model and a previously developed DVH prediction model for lung/mediastinum IMRT planning are used as the information centers within a directed e-learning system guided by scoring criteria and communicated with the trainees via a user interface ran from the treatment planning system (Eclipse). The scoring system serves both to illustrate relative quality of plans and to serve as a guide to facilitate directed changes within the plan. One patient serves as a benchmark to show skill development from the e-learning system and is completed without intervention. Five additional lung/mediastinum patients follow in the subsequent training pipeline where the models, graphical user interface (GUI) and trainer work with trainee’s directives and guide meaningful beam selection and tradeoffs within IMRT optimization. Five trainees with minimal treatment planning background were evaluated by both the scoring criteria and a physician to look for improved planning quality and relative effectiveness against the clinically delivered plan.

Trainees scored an average of 22.7% of the total points within the scoring criteria for their benchmark yet improved to an average of 51.9% compared to the clinically delivered plan which achieved 54.1% of the total potential points. Two of the five trainee final plans were rated as comparable to the clinically delivered by a physician and all five were noticeably improved by the physicians standards. For plans within the system, trainees performed on average 24.5% better than the clinically delivered plan with respect to the scoring criteria.

This first attempt at creating a dynamic interface communicating prior experience built in models to an end-user was approximately 10 hours to rapidly improve planning quality. It brings unexperienced planners to a level comparable of experienced dosimetrists for a specific treatment site and when used to inform decisions, the knowledge-based models aided in producing high quality plans.