Evaluation and scoring of radiotherapy treatment plans using an artificial neural network

Predicting treatment plan approval probability for high-dose-rate brachytherapy of cervical cancer using adversarial deep learning

Objective.Predicting the probability of having the plan approved by the physician is important for automatic treatment planning. Driven by the mathematical foundation of deep learning that can use a deep neural network to represent functions accurately and flexibly, we developed a deep-learning framework that learns the probability of plan approval for cervical cancer high-dose-rate brachytherapy (HDRBT).Approach.The system consisted of a dose prediction network (DPN) and a plan-approval probability network (PPN). DPN predicts organs at risk (OAR)D2ccand CTVD90%of the current fraction from the patient's current anatomy and prescription dose of HDRBT. PPN outputs the probability of a given plan being acceptable to the physician based on the patients anatomy and the total dose combining HDRBT and external beam radiotherapy sessions. Training of the networks was achieved by first training them separately for a good initialization, and then jointly via an adversarial process. We collected approved treatment plans of 248 treatment fractions from 63 patients. Among them, 216 plans from 54 patients were employed in a four-fold cross validation study, and the remaining 32 plans from other 9 patients were saved for independent testing.Main results.DPN predicted equivalent dose of 2 Gy for bladder, rectum, sigmoidD2ccand CTVD90%with a relative error of 11.51% ± 6.92%, 8.23% ± 5.75%, 7.12% ± 6.00%, and 10.16% ± 10.42%, respectively. In a task that differentiates clinically approved plans and disapproved plans generated by perturbing doses in ground truth approved plans by 20%, PPN achieved accuracy, sensitivity, specificity, and area under the curve 0.70, 0.74, 0.65, and 0.74.Significance.We demonstrated the feasibility of developing a novel deep-learning framework that predicts a probability of plan approval for HDRBT of cervical cancer, which is an essential component in automatic treatment planning.

Modeling physician's preference in treatment plan approval of stereotactic body radiation therapy of prostate cancer

Objective.Treatment planning of radiation therapy is a time-consuming task. It is desirable to develop automatic planning approaches to generate plans favorable to physicians. The purpose of this study is to develop a deep learning based virtual physician network (VPN) that models physician's preference on plan approval for prostate cancer stereotactic body radiation therapy (SBRT).Approach.VPN takes one planning target volume (PTV) and eight organs at risk structure images, as well as a dose distribution of a plan seeking approval as input. It outputs a probability of approving the plan, and a dose distribution indicating improvements to the input dose. Due to the lack of unapproved plans in our database, VPN is trained using an adversarial framework. 68 prostate cancer patients who received 45Gyin 5-fraction SBRT were selected in this study, with 60 patients for training and cross validation, and 8 patients for independent testing.Main results.The trained VPN was able to differentiate approved and unapproved plans with Area under the curve 0.97 for testing data. For unapproved plans, after applying VPN's suggested dose improvement, the improved dose agreed with ground truth with relative differences2.03±2.17%for PTVD98%,0.49±0.29%for PTVV95%,3.08±2.24%for penile bulbDmean,3.73±2.20%for rectumV50%,and2.06±1.73%for bladderV50%.Significance.VPN was developed to accurately model a physician's preference on plan approval and to provide suggestions on how to improve the dose distribution.

Machine learning applications in radiation oncology

Machine learning technology has a growing impact on radiation oncology with an increasing presence in research and industry. The prevalence of diverse data including 3D imaging and the 3D radiation dose delivery presents potential for future automation and scope for treatment improvements for cancer patients. Harnessing this potential requires standardization of tools and data, and focused collaboration between fields of expertise. The rapid advancement of radiation oncology treatment technologies presents opportunities for machine learning integration with investments targeted towards data quality, data extraction, software, and engagement with clinical expertise. In this review, we provide an overview of machine learning concepts before reviewing advances in applying machine learning to radiation oncology and integrating these techniques into the radiation oncology workflows. Several key areas are outlined in the radiation oncology workflow where machine learning has been applied and where it can have a significant impact in terms of efficiency, consistency in treatment and overall treatment outcomes. This review highlights that machine learning has key early applications in radiation oncology due to the repetitive nature of many tasks that also currently have human review. Standardized data management of routinely collected imaging and radiation dose data are also highlighted as enabling engagement in research utilizing machine learning and the ability integrate these technologies into clinical workflow to benefit patients. Physicists need to be part of the conversation to facilitate this technical integration.

Artificial Intelligence: reshaping the practice of radiological sciences in the 21st century

Advances in computing hardware and software platforms have led to the recent resurgence in artificial intelligence (AI) touching almost every aspect of our daily lives by its capability for automating complex tasks or providing superior predictive analytics. AI applications are currently spanning many diverse fields from economics to entertainment, to manufacturing, as well as medicine. Since modern AI's inception decades ago, practitioners in radiological sciences have been pioneering its development and implementation in medicine, particularly in areas related to diagnostic imaging and therapy. In this anniversary article, we embark on a journey to reflect on the learned lessons from past AI's chequered history. We further summarize the current status of AI in radiological sciences, highlighting, with examples, its impressive achievements and effect on re-shaping the practice of medical imaging and radiotherapy in the areas of computer-aided detection, diagnosis, prognosis, and decision support. Moving beyond the commercial hype of AI into reality, we discuss the current challenges to overcome, for AI to achieve its promised hope of providing better precision healthcare for each patient while reducing cost burden on their families and the society at large.

Machine learning and modeling: Data, validation, communication challenges

With the era of big data, the utilization of machine learning algorithms in radiation oncology is rapidly growing with applications including: treatment response modeling, treatment planning, contouring, organ segmentation, image-guidance, motion tracking, quality assurance, and more. Despite this interest, practical clinical implementation of machine learning as part of the day-to-day clinical operations is still lagging. The aim of this white paper is to further promote progress in this new field of machine learning in radiation oncology by highlighting its untapped advantages and potentials for clinical advancement, while also presenting current challenges and open questions for future research. The targeted audience of this paper includes newcomers as well as practitioners in the field of medical physics/radiation oncology. The paper also provides general recommendations to avoid common pitfalls when applying these powerful data analytic tools to medical physics and radiation oncology problems and suggests some guidelines for transparent and informative reporting of machine learning results.

Artificial neural network based gynaecological image-guided adaptive brachytherapy treatment planning correction of intra-fractional organs at risk dose variation

PURPOSE

Intra-fractional organs at risk (OARs) deformations can lead to dose variation during image-guided adaptive brachytherapy (IGABT). The aim of this study was to modify the final accepted brachytherapy treatment plan to dosimetrically compensate for these intra-fractional organs-applicators position variations and, at the same time, fulfilling the dosimetric criteria.

MATERIAL AND METHODS

Thirty patients with locally advanced cervical cancer, after external beam radiotherapy (EBRT) of 45-50 Gy over five to six weeks with concomitant weekly chemotherapy, and qualified for intracavitary high-dose-rate (HDR) brachytherapy with tandem-ovoid applicators were selected for this study. Second computed tomography scan was done for each patient after finishing brachytherapy treatment with applicators in situ. Artificial neural networks (ANNs) based models were used to predict intra-fractional OARs dose-volume histogram parameters variations and propose a new final plan.

RESULTS

A model was developed to estimate the intra-fractional organs dose variations during gynaecological intracavitary brachytherapy. Also, ANNs were used to modify the final brachytherapy treatment plan to compensate dosimetrically for changes in 'organs-applicators', while maintaining target dose at the original level.

CONCLUSIONS

There are semi-automatic and fast responding models that can be used in the routine clinical workflow to reduce individually IGABT uncertainties. These models can be more validated by more patients' plans to be able to serve as a clinical tool.

Quantitative metrics for assessing plan quality

Despite many studies over the last 3 decades that have attempted to explicitly quantify the decision-making process for radiotherapy treatment plan evaluation, judgments of an individual plan's degree of quality are still largely subjective and can show inter- and intra-practitioner variability even if the clinical treatment goals are the same. Several factors conspire to confound the full quantification of treatment plan quality, including uncertainties in dose response of cancerous and normal tissue, the rapid pace of new technology adoption, and the human component of treatment planning. However, new developments in clinical informatics and automation are lowering the bar for developing and implementing quantitative metrics into the treatment planning process. This review discusses general strategies for using quantitative metrics in the treatment planning process and presents a case study in intensity-modulated radiation therapy planning whereby control was established on a variable system via such techniques.

Application of the Naïve Bayesian Classifier to optimize treatment decisions

BACKGROUND AND PURPOSE

To study the accuracy, specificity and sensitivity of the Naïve Bayesian Classifier (NBC) in the assessment of individual risk of cancer relapse or progression after radiotherapy (RT).

MATERIALS AND METHODS

Data of 142 brain tumour patients irradiated from 2000 to 2005 were analyzed. Ninety-six attributes related to disease, patient and treatment were chosen. Attributes in binary form consisted of the training set for NBC learning. NBC calculated an individual conditional probability of being assigned to: relapse or progression (1), or no relapse or progression (0) group. Accuracy, attribute selection and quality of classifier were determined by comparison with actual treatment results, leave-one-out and cross validation methods, respectively. Clinical setting test utilized data of 35 patients. Treatment results at classification were unknown and were compared with classification results after 3 months.

RESULTS

High classification accuracy (84%), specificity (0.87) and sensitivity (0.80) were achieved, both for classifier training and in progressive clinical evaluation.

CONCLUSIONS

NBC is a useful tool to support the assessment of individual risk of relapse or progression in patients diagnosed with brain tumour undergoing RT postoperatively.

Interactively exploring optimized treatment plans

PURPOSE

A new paradigm for treatment planning is proposed that embodies the concept of interactively exploring the space of optimized plans. In this approach, treatment planning ignores the details of individual plans and instead presents the physician with clinical summaries of sets of solutions to well-defined clinical goals in which every solution has been optimized in advance by computer algorithms.

METHODS AND MATERIALS

Before interactive planning, sets of optimized plans are created for a variety of treatment delivery options and critical structure dose-volume constraints. Then, the dose-volume parameters of the optimized plans are fit to linear functions. These linear functions are used to show in real time how the target dose-volume histogram (DVH) changes as the DVHs of the critical structures are changed interactively. A bitmap of the space of optimized plans is used to restrict the feasible solutions. The physician selects the critical structure dose-volume constraints that give the desired dose to the planning target volume (PTV) and then those constraints are used to create the corresponding optimized plan.

RESULTS

The method is demonstrated using prototype software, Treatment Plan Explorer (TPEx), and a clinical example of a patient with a tumor in the right lung. For this example, the delivery options included 4 open beams, 12 open beams, 4 wedged beams, and 12 wedged beams. Beam directions and relative weights were optimized for a range of critical structure dose-volume constraints for the lungs and esophagus. Cord dose was restricted to 45 Gy. Using the interactive interface, the physician explored how the tumor dose changed as critical structure dose-volume constraints were tightened or relaxed and selected the best compromise for each delivery option. The corresponding treatment plans were calculated and compared with the linear parameterization presented to the physician in TPEx. The linear fits were best for the maximum PTV dose and worst for the minimum PTV dose. Based on the root-mean-square error between the fit values and their corresponding data values, the linear fit appears to be adequate, although higher order polynomials could give better results. Some of the variance in fit is due to the stochastic nature of the simulated annealing optimization algorithm, which does not reproduce the exact same results in repetitions of the same calculation. Using a directed search algorithm for plan optimization should produce better parameter fits and, therefore, better predictions of plan characteristics by TPEx.

CONCLUSIONS

Using TPEx, the physician can easily select the optimum plan for a patient, with no imposed arbitrary definition of the "best" plan. More importantly, the physician can readily see what can be achieved for the patient with a given delivery technique. There is no more uncertainty about whether or not a better plan exists. By comparing the "best" plans for different delivery options (e.g., three-dimensional conformal radiotherapy versus intensity-modulated radiation therapy), the physician can gauge the clinical benefits of greater technical complexity. However, before the TPEx process can be clinical useful, faster computers and/or algorithms are needed and more studies are needed to better model the spaces of optimized solutions.

Inverse treatment planning using volume-based objective functions

The results of optimization of inverse treatment plans depend on a choice of the objective function. Even when the optimal solution for a given cost function can be obtained, a better solution may exist for a given clinical scenario and it could be obtained with a revised objective function. In the approach presented in this work mixed integer programming was used to introduce a new volume-based objective function, which allowed for minimization of the number of under- or overdosed voxels in selected structures. By selecting and prioritizing components of this function the user could drive the computations towards the desired solution. This optimization approach was tested using cases of patients treated for prostate and oropharyngeal cancer. Initial solutions were obtained based on minimization/maximization of the dose to critical structures and targets. Subsequently, the volume-based objective functions were used to locate solutions, which satisfied better clinical objectives particular to each of the cases. For prostate cases, these additional solutions offered further improvements in sparing of the rectum or the bladder. For oropharyngeal cases, families of solutions were obtained satisfying an intensity modulated radiation therapy protocol for this disease site, while offering significant improvement in the sparing of selected critical structures, e.g., parotid glands. An additional advantage of the present approach was in providing a convenient mechanism to test the feasibility of the dose-volume histogram constraints.

An integrated two-level hierarchical system for decision making in radiation therapy based on fuzzy cognitive maps

The radiation therapy decision-making is a complex process that has to take into consideration a variety of interrelated functions. Many fuzzy factors that must be considered in the calculation of the appropriate dose increase the complexity of the decision-making problem. A novel approach introduces fuzzy cognitive maps (FCMs) as the computational modeling method, which tackles the complexity and allows the analysis and simulation of the clinical radiation procedure. Specifically this approach is used to determine the success of radiation therapy process estimating the final dose delivered to the target volume, based on the soft computing technique of FCMs. Furthermore a two-level integrated hierarchical structure is proposed to supervise and evaluate the radiotherapy process prior to treatment execution. The supervisor determines the treatment variables of cancer therapy and the acceptance level of final radiation dose to the target volume. Two clinical case studies are used to test the proposed methodology and evaluate the simulation results. The usefulness of this two-level hierarchical structure discussed and future research directions are suggested for the clinical use of this methodology.

Generating compensation designs for tangential breast irradiation with artificial neural networks

In this paper we discuss a study comparing an algorithm implemented clinically to design intensity-modulated fields with two artificial neural networks (ANNs) trained to design the same fields. The purpose of the algorithm is to produce compensation for tangential breast radiotherapy in order to improve dose homogeneity. This was achieved by creating intensity-modulated fields to supplement standard wedged fields. Portal image data were used to create thickness maps of the medial and lateral fields, which in turn were used to design the wedged and intensity-modulated fields. The ANNs were developed to design the intensity-modulated fields from the portal image data and corresponding fluence map alone. One used localized groups of portal image pixels related to the fluence map (method 2), and the other used a one-to-one mapping between spatially corresponding pixels (method 3). A dosimetric comparison of the methods was performed by calculating the overall dose distribution. The volume of tissue outside the dose range 95-105% was used to assess dose homogeneity. The average volume outside 95-105%, averaged over 80 cases, was shown to be 2.3% for the algorithm, whilst average values of 9.9% and 13.5% were obtained for methods 2 and 3, respectively. The results of this study demonstrate the ability of an ANN to learn the general shape of compensation required and explore the use of image-based ANNs in the design of intensity-modulated fields.

Treatment planning using a dose-volume feasibility search algorithm

PURPOSE

An approach to treatment plan optimization is presented that inputs dose--volume constraints and utilizes a feasibility search algorithm that seeks a set of beam weights so that the calculated dose distributions satisfy the dose--volume constraints. In contrast to a search for the "best" plan, this approach can quickly determine feasibility and point out the most restrictive of the predetermined constraints.

METHODS AND MATERIALS

The cyclic subgradient projection (CSP) algorithm was modified to incorporate dose--volume constraints in a treatment plan optimization schema. The algorithm was applied to determine beam weights for several representative three-dimensional treatment plans.

RESULTS

Using the modified CSP algorithm, we found that either a feasible solution to the dose--volume constraint problem was found or the program determined, after a predetermined set of iterations was performed, that no feasible solution existed for the particular set of dose--volume constraints. If no feasible solution existed, we relaxed several of the dose--volume constraints and were able to achieve a feasible solution.

CONCLUSION

Feasibility search algorithms can be used in radiation treatment planning to generate a treatment plan that meets the dose--volume constraints established by the radiation oncologist. In the absence of a feasible solution, these algorithms can provide information to the radiation oncologist as to how the dose--volume constraints may be modified to achieve a feasible solution.

Beam-orientation customization using an artificial neural network

A methodology for the constrained customization of coplanar beam orientations in radiotherapy treatment planning using an artificial neural network (ANN) has been developed. The geometry of the patients, with cancer of the prostate, was modelled by reducing the external contour, planning target volume (PTV) and organs at risk (OARs) to a set of cuboids. The coordinates and size of the cuboids were given to the ANN as inputs. A previously developed beam-orientation constrained-customization (BOCC) scheme employing a conventional computer algorithm was used to determine the customized beam orientations in a training set containing 45 patient datasets. Twelve patient datasets not involved in the training of the artificial neural network were used to test whether the ANN was able to map the inputs to customized beam orientations. Improvements from the customized beam orientations were compared with standard treatment plans with fixed gantry angles and plans produced from the BOCC scheme. The ANN produced customized beam orientations within 5 degrees of the BOCC scheme in 62.5% of cases. The average difference in the beam orientations produced by the ANN and the BOCC scheme was 7.7 degrees (+/-1.7, 1 SD). Compared with the standard treatment plans, the BOCC scheme produced plans with an increase in the average tumour control probability (TCP) of 5.7% (+/-1.4, 1 SD) whilst the ANN generated plans increased the average TCP by 3.9% (+/-1.3, 1 SD). Both figures refer to the TCP at a fixed rectal normal tissue complication probability (NTCP) of 1%. In conclusion, even using a very simple model for the geometry of the patient, an ANN was able to produce beam orientations that were similar to those produced by a conventional computer algorithm.

A neural network to predict symptomatic lung injury

A nonlinear neural network that simultaneously uses pre-radiotherapy (RT) biological and physical data was developed to predict symptomatic lung injury. The input data were pre-RT pulmonary function, three-dimensional treatment plan doses and demographics. The output was a single value between 0 (asymptomatic) and 1 (symptomatic) to predict the likelihood that a particular patient would become symptomatic. The network was trained on data from 97 patients for 400 iterations with the goal to minimize the mean-squared error. Statistical analysis was performed on the resulting network to determine the model's accuracy. Results from the neural network were compared with those given by traditional linear discriminate analysis and the dose-volume histogram reduction (DVHR) scheme of Kutcher. Receiver-operator characteristic (ROC) analysis was performed on the resulting network which had Az = 0.833 +/- 0.04. (Az is the area under the ROC curve.) Linear discriminate multivariate analysis yielded an Az = 0.813 +/- 0.06. The DVHR method had Az = 0.521 +/- 0.08. The network was also used to rank the significance of the input variables. Future studies will be conducted to improve network accuracy and to include functional imaging data.

Artificial neural network model of survival in patients treated with irradiation with and without concurrent chemotherapy for advanced carcinoma of the head and neck

PURPOSE

This study was performed to investigate the feasibility of predicting survival in squamous cell carcinoma of the head and neck (SCCHN) with an artificial neural network (ANN), and to compare ANN performance with conventional models.

METHODS AND MATERIALS

Data were analyzed from a Phase III trial in which patients with locally advanced SCCHN received hyperfractionated irradiation with or without concurrent cisplatin and 5-fluorouracil. Of the 116 randomized patients, 95 who had 2-year follow-up and all required data were evaluated. ANN and logistic regression (LR) models were constructed to predict 2-year total survival using round-robin cross-validation. A modified staging model was also examined.

RESULTS

The best LR model used tumor size, nodal stage, and race to predict survival. The best ANN used nodal stage, tumor size, stage, and resectability, and hemoglobin. Treatment type did not predict 2-year survival and was not included in either model. Using the respective best feature sets, the area under the receiver operating characteristic curve (Az) for the ANN was 0.78 +/- 0.05, showing more accurate overall performance than LR (Az = 0.67 +/- 0.05, p = 0.07). At 70% sensitivity, the ANN was 72% specific, while LR was 54% specific (p = 0.08). At 70% specificity, the ANN was 72% sensitive, while LR was 54% sensitive (p = 0.07). When both models used the five predictive variables best for an ANN, Az for LR decreased [Az = 0.61 +/- 0.06, p < 0.01 (ANN)]. The models performed equivalently when using the three variables best for LR. The best ANN also compared favorably with staging [Az = 0.60 +/- 0.07, p = 0.02 (ANN)].

CONCLUSIONS

An ANN modeled 2-year survival in this data set more accurately than LR or staging models and employed predictive variables that could not be used by LR. Further work is planned to confirm these results on larger patient samples, examining longer follow-up to incorporate treatment type into the model.

Using an artificial neural network to define the planning target volume in radiotherapy

A neural network for predicting the planning target volume in radiotherapy from the shape of the detected tumor is designed and tested in this research project. The proposed neural network is able to generalize expert medical knowledge and predict the planning target volume from a three-dimensional image of the detected tumor. Initial results for simple shaped brain tumors are presented in this paper.