Hao H. Zhang

Moderate predictive value of demographic and behavioral characteristics for a diagnosis of HPV16-positive and HPV16-negative head and neck cancer

Patients with HPV-positive and HPV-negative head and neck squamous cell carcinoma (HNSCC) are significantly different with regard to sociodemographic and behavioral characteristics that clinicians may use to assume tumor HPV status. Machine learning methods were used to evaluate the predictive value of patient characteristics and laboratory biomarkers of HPV exposure for a diagnosis of HPV16-positive HNSCC compared to in situ hybridization, the current gold-standard. Models that used a combination of demographic characteristics such as age, tobacco use, gender, and race had only moderate predictive value for tumor HPV status among all patients with HNSCC (positive predictive value [PPV]=75%, negative predictive value [NPV]=68%) or when limited to oropharynx cancer patients (PPV=55%, NPV=65%) and thus included a sizeable number of false positive and false negative predictions. Prediction was not improved by the addition of other demographic or behavioral factors (sexual behavior, income, education) or biomarkers of HPV16 exposure (L1, E6/7 antibodies or DNA in oral exfoliated cells). Patient demographic and behavioral characteristics as well as HPV biomarkers are not an accurate substitute for clinical testing of tumor HPV status.

Modeling Plan-Related Clinical Complications Using Machine Learning Tools in a Multiplan IMRT Framework

PURPOSE To predict organ-at-risk (OAR) complications as a function of dose-volume (DV) constraint settings without explicit plan computation in a multiplan intensity-modulated radiotherapy (IMRT) framework. METHODS AND MATERIALS Several plans were generated by varying the DV constraints (input features) on the OARs (multiplan framework), and the DV levels achieved by the OARs in the plans (plan properties) were modeled as a function of the imposed DV constraint settings. OAR complications were then predicted for each of the plans by using the imposed DV constraints alone (features) or in combination with modeled DV levels (plan properties) as input to machine learning (ML) algorithms. These ML approaches were used to model two OAR complications after head-and-neck and prostate IMRT: xerostomia, and Grade 2 rectal bleeding. Two-fold cross-validation was used for model verification and mean errors are reported. RESULTS Errors for modeling the achieved DV values as a function of constraint settings were 0-6%. In the head-and-neck case, the mean absolute prediction error of the saliva flow rate normalized to the pretreatment saliva flow rate was 0.42% with a 95% confidence interval of (0.41-0.43%). In the prostate case, an average prediction accuracy of 97.04% with a 95% confidence interval of (96.67-97.41%) was achieved for Grade 2 rectal bleeding complications. CONCLUSIONS ML can be used for predicting OAR complications during treatment planning allowing for alternative DV constraint settings to be assessed within the planning framework.

The minimum knowledge base for predicting organ-at-risk dose–volume levels and plan-related complications in IMRT planning

IMRT treatment planning requires consideration of two competing objectives: achieving the required amount of radiation for the planning target volume and minimizing the amount of radiation delivered to all other tissues. It is important for planners to understand the tradeoff between competing factors so that the time-consuming human interaction loop (plan-evaluate-modify) can be eliminated. Treatment-plan-surface models have been proposed as a decision support tool to aid treatment planners and clinicians in choosing between rival treatment plans in a multi-plan environment. In this paper, an empirical approach is introduced to determine the minimum number of treatment plans (minimum knowledge base) required to build accurate representations of the IMRT plan surface in order to predict organ-at-risk (OAR) dose-volume (DV) levels and complications as a function of input DV constraint settings corresponding to all involved OARs in the plan. We have tested our approach on five head and neck patients and five whole pelvis/prostate patients. Our results suggest that approximately 30 plans were sufficient to predict DV levels with less than 3% relative error in both head and neck and whole pelvis/prostate cases. In addition, approximately 30-60 plans were sufficient to predict saliva flow rate with less than 2% relative error and to classify rectal bleeding with an accuracy of 90%.

A two-stage sequential linear programming approach to IMRT dose optimization

The conventional IMRT planning process involves two stages in which the first stage consists of fast but approximate idealized pencil beam dose calculations and dose optimization and the second stage consists of discretization of the intensity maps followed by intensity map segmentation and a more accurate final dose calculation corresponding to physical beam apertures. Consequently, there can be differences between the presumed dose distribution corresponding to pencil beam calculations and optimization and a more accurately computed dose distribution corresponding to beam segments that takes into account collimator-specific effects. IMRT optimization is computationally expensive and has therefore led to the use of heuristic (e.g., simulated annealing and genetic algorithms) approaches that do not encompass a global view of the solution space. We modify the traditional two-stage IMRT optimization process by augmenting the second stage via an accurate Monte Carlo-based kernel-superposition dose calculations corresponding to beam apertures combined with an exact mathematical programming-based sequential optimization approach that uses linear programming (SLP). Our approach was tested on three challenging clinical test cases with multileaf collimator constraints corresponding to two vendors. We compared our approach to the conventional IMRT planning approach, a direct-aperture approach and a segment weight optimization approach. Our results in all three cases indicate that the SLP approach outperformed the other approaches, achieving superior critical structure sparing. Convergence of our approach is also demonstrated. Finally, our approach has also been integrated with a commercial treatment planning system and may be utilized clinically.

SU‐E‐T‐611: Utilizing Machine Learning Techniques for Beam Angle Selection in Radiation Treatment Planning

PURPOSE To utilize machine learning techniques within beam angle optimization to determine an optimal Intensity-modulated radiation therapy (IMRT) beam angle set. METHODS The input data were derived from a collection of equally-spaced seven-beam plans (e-plans) generated using the Pinnacle. This collection of e-plans contains all 72 beam angles corresponding to 5 degree spacing, and the dose delivered to patient tissues from each of these 72 angles was extracted to generate p-scores. Equally-spaced beam sets are commonly used in clinical practice, so this set of plans not only provides initial input data for our beam angle selection (BAS) procedure, but also provides a good set of benchmarks against which treatment improvement may be measured. A beam set scoring function was developed based on a weighted sum of overdose/underdose criteria. The Nested Partitions (NP) global optimization framework is then utilized to guide a sample-based search for the global optimal of the beam angle space. In our NP-based approach to BAS, a single sample is a 7-beam set satisfying beam spacing constraints. A fast scoring method based on the e-plan single-beam dose data was used to obtain an initial approximate score (c-score) and a set of dose component scores for each beam set. Machine learning techniques were then employed to predict each dose component, and these values were used to compute a predicted score. RESULTS The average improvements in p-scores for 5 cases were 43%, 29% and 11% comparing to default eplan, best eplan and conventional NP (without ML). 10%, 12% and 15% improvement was achieved for sparing of spinal cord, brain stem and oral mucosa, respectively. CONCLUSIONS Machine learning tools provide an effective technique for rapid high-quality approximate scoring for beam angle sets in IMRT. This approximation process leads to excellent beam sets when embedded within the NP global optimization framework. This work was supported in part by a grant from the NIH/NCI CA130814.

Machine Learning for the Prediction of Organ Damage from Cancer Radiotherapy

Abstract There are over one million cases of cancer diagnosed each year just in the United States, and many times that number in other countries. About 60% of US cancer patients are treated with radiotherapy, and increasingly complex radiation delivery procedures are being developed in order to improve treatment outcomes. A key goal is to determine appropriate values for a large set of delivery parameters in order to ensure that as large a fraction as possible of the radiation that enters the patient is delivered to the tumor as opposed to depositing it in adjacent non-cancerous organs that can be damaged by radiation (the latter are termed organs-at-risk (OARs)). In the research described here, we address the issue of tissue damage by developing machine learning (ML) frameworks for the prediction of two types of possible OAR complications associated with an advanced form of treatment known as Intensity Modulated Radiation Therapy. Specifically, we show that ML techniques may be used to develop models for the prediction of radiation-induced (1) xerostomia (“dry mouth”, a significant quality-of-life concern associated with radiation damage to the parotid glands) and (2) rectal bleeding. These patient-specific ML models provide accurate complication prediction on the basis of the selected clinician settings for the radiation constraints. This research will thus allow clinicians to determine treatment parameter values that result in appropriate radiation levels for critical organs.

MO‐D‐351‐03: A Unified Approach to Beam Angle Selection and Dose Optimization with High‐Throughput Computing for IMRT

Purpose: To present a unified approach to solving the Beam Angle Selection (BAS) and DoseOptimization (DO) problems in radiation treatment planning using a Nested Partitions (NP) framework. Method and Materials: The NP framework is a powerful new optimization paradigm that combines adaptive global sampling with local heuristic search. It uses a flexible partitioning method to divide the search space into regions that can be analyzed individually and then coordinates the results to determine how to continue the search, that is, where to concentrate additional computational effort. This partitioning/sampling approach makes the NP framework uniquely well‐suited for high‐throughput computing. Beam angle space is partitioned and sampled. DO algorithms are incorporated during the evaluation of the quality of a selected angle set. After the execution of our proposed method, we not only obtain a set of beam, but also the optimized intensity for each beam in 3DCRT or “intensity maps” for each angle in IMRT.Results: Using a 3DCRT data set of a pancreas case, we demonstrated the following improvements in OAR dose relative to a equispaced beam set: cord, 66%; kidney, 78%; liver 36%. We also considered an IMRT head‐and‐neck case, and obtained a 28% reduction in dose to normal tissue as well as improvements in the right parotid dose with no significant changes in dose to other tissues.Conclusion: We have demonstrated that our framework provides an effective and automated approach to obtaining high‐quality solutions to the unified BAS and DO problems in both 3DCRT and IMRT. Relative to beam‐angle sets constructed via expert clinical judgment and other approaches, the beams and doses via NP with HTC showed significant reduction in the radiation delivered to non‐cancerous tissue near the tumors.

SU‐FF‐T‐242: High‐Throughput Computing in Condor Using a Nested Partitions Framework for IMRT Beam and Dose Optimization

Purpose: To utilize a High Throughput Computing (HTC) system to provide distributed computing on a network of computer workstations for a metaheuristic, Nested Partitions (NP), for beam angle selection (BAS) in IMRT.Method and Materials: An important feature of the NP approach is that its partition/sampling processes are naturally parallelized , allowing the integration of this powerful optimization methodology with high‐performance distributed environments such as the Condor system. Condor is a specialized workload management system that can used on existing computer networks for compute‐intensive jobs, allowing the submission of many jobs at the same time, and performing these jobs by locating and utilizing idle computers in the network. Our algorithm first solves an integer programming formulation of the BAS problem that uses aggregate mean organ‐at‐risk data. Based on this initial angle set, NP uses sampling on a partition of the global solution space to generate alternative angle sets that are distributed via Condor for evaluation via dose optimization. The corresponding dose optimization problems for the beam set samples are thus solved in parallel. NP repeats this procedure iteratively to obtain high quality angle sets for IMRT.Results: Compared to evaluating each beam angle set sequentially, parallel evaluation of the NP‐selected beam angle sets 2457_2via Condor on 15 workstations yielded a factor of 14 speedup. Overall, BAS for a difficult clinical case in this setting required 15–28 minutes. Using a weighted sum of percentage violations of DVH thresholds for OARs as an overall measure of solution quality, a 78% improvement (head‐and‐neck case) and a 30% improvement (pelvic case) were achieved over equally‐spaced beams. Conclusions: High‐quality beam angle sets for IMRT can be efficiently obtained in an automated manner by utilizing an HTC system on an existing computer network to implement an NP‐based metaheuristic. Conflict of Interest: N/A.

MO‐D‐M100J‐07: Plan Space Modeling and Decision Support System for Multi‐Plan IMRT Framework

Purpose: The conventional treatment planning paradigm involves a plan‐and‐evaluate followed by a modify‐if‐necessary approach. We describe an IMRTtreatment planning framework in which multiple plans that differ in the input constraints are generated per case prior to evaluation. We also describe a decision‐support‐system (DSS) for evaluating and ranking multiple plans and hypothesize that the planning surface can be modeled using quadratic modeling. Methods and Materials: One hundred twenty‐five plans were generated sequentially for a head‐and‐neck case and a pelvic case by varying the dose‐volume constraints on each of the OARs. A DSS was used to rank plans according to DVH and equivalent uniform dose (EUD) values using composite criteria and pre‐emptive selection. Two methods for ranking treatment plans were evaluated: composite criteria and pre‐emptive selection. The planning surfaces corresponding to the 125‐plan sets were modeled using quadratic functions by formulating the problem as a linear program. Results: The DSS provides an interface for the comparison of multiple plan features. Plan ranking using both composite and pre‐emptive criteria resulted in the reduction of plan space to 1–3 “optimal” plans. The planning surface models had good predictive capability with respect to both DVH and EUD values with fit errors of < 6%. Models generated by minimizing the maximum relative error had significantly lower relative errors than models obtained by minimizing the sum of squared errors. The inter‐dependence of OAR‐OAR properties could be successfully inferred for both clinical cases through the use of contour plots, which represent projections through the multi‐dimensional plan property space. Conclusion: The DSS can be used to aid the planner in the selection of the most desirable plan. The collection of quadratic models constructed from the plan data in order to predict DVH and EUD values generally showed excellent agreement with the actual plan values.

SU‐FF‐T‐30: A Nested Partitions Framework for Beam Angle and Dose Optimization in IMRT

Purpose: To present a novel algorithm, nested partitions (NP), capable of finding suitable beam angle samples for IMRTtreatment planning by guiding the dose optimization process. Beam angle optimization and dose optimization are two problems which are conventionally solved separately, because coupling the variables increases the size and complexity of the combinatorial optimization problem. Method and Materials: NP is a metaheuristic algorithm, guiding the search of a deterministic dose optimization algorithm. The NP method adaptively samples from the entire feasible region, or search space, and concentrates the sampling effort by systematically partitioning the feasible region at successive iterations. We used a “warm‐start” approach by initiating the NP with beam angle samples derived from an integer programming (IP) model. We implemented the NP framework in conjunction with a quasi‐newton dose‐optimization algorithm employed in a commercial treatment planning system. We evaluated the results using 7‐field plans for two test clinical cases: head and neck and pancreas. Results: The results of four iterations of the NP algorithm outperformed both the initial IP solution and a generic equi‐spaced beam angle plan. This evaluation was based on DVH constraints for the critical structures for both clinical cases. For example, in the head and neck case, the NP plan delivered a dose of greater than 35 Gy to just 4.3% of the spinal cord, compared to 5.2% for the IP plan and 41.4% for the generic plan. In the pancreas case, the NP delivered a dose of greater than 23 Gy to 30.9% of the right kidney, compared to 43.7% (IP plan) and 49.0% (generic plan). Conclusions: Our results indicate that the IP solution provides a good initial solution. In addition, by employing the NP framework, further improvement is achieved. This makes it possible to produce a high‐quality solution within a reasonable amount of time.