This protocol describes a series of automated tools designed for high-quality radiotherapy autocontouring and autoplanning that are being packaged into a web-based service to maximize robustness and scalability while minimizing operational costs.
Access to radiotherapy worldwide is limited. The Radiation Planning Assistant (RPA) is a fully automated, web-based tool that is being developed to offer fully automated radiotherapy treatment planning tools to clinics with limited resources. The goal is to help clinical teams scale their efforts, thus reaching more patients with cancer. The user connects to the RPA via a webpage, completes a Service Request (prescription and information about the radiotherapy targets), and uploads the patient's CT image set. The RPA offers two approaches to automated planning. In one-step planning, the system uses the Service Request and CT scan to automatically generate the necessary contours and treatment plan. In two-step planning, the user reviews and edits the automatically generated contours before the RPA continues to generate a volume-modulated arc therapy plan. The final plan is downloaded from the RPA website and imported into the user's local treatment planning system, where the dose is recalculated for the locally commissioned linac; if necessary, the plan is edited prior to approval for clinical use.
The global number of cancer cases is expected to grow to approximately 24.6 million by 2030, with the largest burden in low- and middle-income countries (LMICs)1. Radiotherapy is a cost-effective, curative, and palliative treatment for cancer, providing benefits to approximately 50% of patients with cancer and to 60-70% in low-income countries where patients are more likely to present at a late stage2,3. However, access to radiotherapy worldwide is limited4; for example, no countries in Africa have the radiotherapy capacity to meet their estimated need5. Several studies have estimated these impending shortages and what would be required to meet upcoming needs6,7.
The Lancet Oncology Commission made the compelling case that investment in radiotherapy capacity improvements would not only save lives but would also bring positive economic benefits3. They also specifically pointed out that automation of radiotherapy contouring and treatment planning may help clinical teams scale their efforts by significantly reducing the time spent on these tasks by oncologists and physicists, respectively, making the goals more attainable.
Our research group has been working in collaboration with clinical teams at MD Anderson and at hospitals across the world to develop web-based automated tools. This suite of tools (called the RPA) provides Artificial Intelligence-based contouring (outlining tumors and nearby organs on CT scans) and radiotherapy treatment planning (which defines exactly how the radiation is delivered). This web-based platform offers the advantage of reduced time and resources needed to prepare high-quality plans for each patient.
Our experience with an early version of an AI-based tool at MD Anderson has shown that automated contouring can save up to 2 h per patient-a significant streamlining of the workflow. This means that current clinical staff will be able to scale their effort, treating more patients with higher-quality radiotherapy. By offering these tools via a fully automated, web-based service (Radiation Planning Assistant [RPA], RPA.mdanderson.org), we can minimize the cost to the patients and providers and maximize the reach of this tool.
We have been developing the RPA for 6 years, and several significant changes have been made since we first published on the RPA workflows8. These include developing the RPA into a web-based tool, thus reducing the costs associated with installation and maintenance, and improving the system's robustness. Other improvements include changes in the user interfaces to improve usability and reduce the risk of error9 and expanding treatment options (specifically, radiotherapy planning for post-mastectomy breast10 and metastases to the brain11). Thus, the protocol described here is substantially more advanced than the initial version previously published.
The RPA uses a one-step process to create contours and plans in situations in which editing of the contours is not generally necessary to create the treatment plan. This includes four-field box treatment planning for cervical cancer (based on bony landmarks or automatically generated soft tissue contours)12,13,14,15, tangential or supraclavicular fields for post-mastectomy breast cancer11, and opposed laterals for whole-brain treatments16. In the near future, we expect to add craniospinal treatments for pediatric cancers17, three-field treatments for rectal cancer18, and treatment planning for various palliative cases (vertebral bodies, hips, and ribs)19, as well as lung and bladder cancers. Currently, more advanced treatments, specifically volume-modulated arc therapy (VMAT), require a two-step process in which automatically generated contours are edited prior to treatment planning13,20. However, the quality of deep learning-based autocontouring is such that we expect to change these planning approaches to a one-step process in the future. This protocol focuses on one-step planning.
Figure 1 shows the overall workflow for creating a radiotherapy treatment plan using the RPA, with more details on the different tasks shown in Table 1. In summary, the RPA requires a completed Service Request (which includes information such as the dose prescription and treatment approach) and a patient's individual CT scan. The Service Request must be accepted by a radiation oncologist. The CT scan must be accepted by a Clinical User to ensure that the RPA calculations are performed on the correct CT scan. Once the RPA has generated a plan, it should be downloaded from the RPA website and imported into the user's treatment planning system, where the dose should be recalculated. This is necessary because the RPA calculates plans on standard beams (available for several linac models), which may not exactly match the beam characteristics of the local linac. This approach was taken to reduce cost, although customization may be needed if the local beams are significantly different from our standard beams. The users (treatment planner and radiation oncologist) may make edits to the plan. The plan then enters the user's typical clinical workflow, including local quality assurance checks. Finally, the user should upload their final (recalculated and edited) plan to the RPA website, where an automated comparison between the Final Plan and the RPA Plan is performed. This is a useful check of data integrity in the overall workflow.
Figure 1: Workflow of the automated treatment planning process. Please click here to view a larger version of this figure.
Task # in Figure 1 | Task description | Location | |||
Review the status of any past patients | Main dashboard | ||||
1 | Complete a Service Request on the RPA website | Service Request dashboard | |||
2 | Upload a CT scan to the RPA website | CT scan dashboard | |||
3 | Check on the status of the patient | Main dashboard | |||
Review and download the RPA Plan | Main dashboard | ||||
4 | Import the plan into the user’s TPS, recalculate the dose, and make edits as necessary | Local TPS | |||
5, 6 | Upload the Final Plan to the RPA website | Plan comparison dashboard | |||
Review the automatic comparison of the Final Plan and the RPA Plan | Plan comparison dashboard | ||||
– | The plan enters the user’s routine clinical workflow, including regular quality assurance | User’s own software |
Table 1: Overview of the tasks involved in creating an RPA Plan using the 1-step workflow. Local TPS: User's Treatment Planning System.
This manuscript describes this one-step workflow for the RPA and presents some example results of the treatment planning process output. Currently, the following planning approaches use this one-step workflow: i) Four-field box treatment plans for cervical cancer patients (bony landmark-based apertures); ii) four-field box treatment plans for cervical cancer patients (soft tissue-based field apertures); iii) tangential and supraclavicular treatment plans for chest wall patients; iv) whole-brain treatment plans.
All patient data used for evaluating the RPA were used retrospectively, with approval from The University of Texas MD Anderson Institutional Review Board. The RPA comprises a series of dashboards located to the left of the main menu of the RPA webpage (Figure 2). Figure 2 shows the main dashboard. All dashboards have a similar look but focus on different tasks and personnel. The following protocol describes the key processes for the automatic creation of a treatment plan.
Figure 2: Screenshot of the RPA main dashboard. Please click here to view a larger version of this figure.
1. Complete Service Request
Status | 概要 | |||
Pending Review | The Service Request for this patient has already been created and is waiting for the radiation oncologist to Accept it. | |||
Accepted | The Service Request for this patient has been Accepted. The status of this patient in the Service Request dashboard will not change until a CT scan for this patient is Accepted. | |||
Rejected by User | The User rejected the Service Request. | |||
Submitted | This case has been submitted to the RPA–more details on this patient’s status can be found on the main dashboard. | |||
RPA System Error | Processing by the RPA was initiated, but the RPA encountered an error and was unable to complete its task. |
Table 2: Patient status categories for the Service Request Dashboard.
Figure 3: An example Service Request created for a soft tissue-based 4-field plan for cervical cancer. Please click here to view a larger version of this figure.
2. Upload CT scan and accept for automated planning
Figure 4 shows a screenshot of the CT review workspace. To upload and review a CT scan:
Figure 4: A screenshot of the CT review workspace. Please click here to view a larger version of this figure.
Status | 概要 | |||
Pending Review | The CT scan has undergone initial processing and is waiting for the user to review and Accept the scan. | |||
Accepted | The CT scan for this patient has been Accepted. Please note that the status of this patient in the CT Scan dashboard will not change until a Service Request is Accepted. | |||
Rejected by User | The CT scan was rejected by the user. | |||
Submitted | This case has been submitted to the RPA–more details on this patient’s status can be found on the main dashboard. | |||
RPA System Error | Processing by the RPA was initiated, but the RPA encountered an error and was unable to complete its tasks. |
Table 3: Patient status categories for the CT Scans dashboard.
3. Monitor planning progress
The status categories of current patients can be viewed on the main dashboard (Table 4). To perform a preliminary review of any completed RPA plan and download it for use:
Status | 概要 | |||
No Accepted CT Scan | This patient has no Accepted CT scan (but an Accepted Service Request is available). | |||
No Accepted Service Request | This patient has no Accepted Service Request (but an Accepted CT scan is available). | |||
Queued | The data for this patient have been sent to the RPA system and are queued for processing. | |||
Processing | Initial processing of this patient’s data is underway. | |||
Processing–contouring | The RPA Contours are being generated. | |||
Processing–contouring report | The RPA Contours Report is being generated. | |||
Complete–RPA contours | Contours generated by the RPA are complete and ready for the user to download and edit (2-step workflows only). | |||
Pending contour review | Planning contours (i.e. contours after the user has made edits/additions) have been uploaded back to the RPA and the Plan Contours Report has been generated. The user needs to Accept this report (from the contours dashboard). | |||
Queued–plan contours | The planning contours (i.e. contours after the user has made edits/additions) are queued for processing prior to the RPA Planning process. | |||
Queued–preplan | This patient’s plan is queued for the planning process. | |||
Processing–preplan | Preplanning processing is underway. | |||
Queued–optimization | This patient’s plan is queued for the plan optimization process. | |||
Processing–optimization | Plan optimization is underway. | |||
Queued–QA | This patient’s plan is queued for the automated quality assurance (QA) process. | |||
Processing–QA | The plan QA is underway. | |||
Processing–plan report | The Final Plan report is being processed. | |||
Complete–RPA Plan | The RPA Plan is complete and ready for download. | |||
Fail–RPA fail | An RPA process has failed. |
Table 4: Patient status categories for the main dashboard.
4. Importing the RPA Plan into the User's treatment planning system and reviewing for clinical use
NOTE: Once the RPA plan is downloaded (DICOM files), the following steps should be completed in the user's treatment planning system:
5. Uploading the Final Plan to the RPA website and reviewing the automatic comparison of the Final Plan and the RPA Plan
Patients for whom an RPA Plan has been generated will appear in the plan comparison dashboard. The plan comparison dashboard provides the status categories shown in Table 5 for current patients. To upload the final User Plan and review an automatic comparison of the User Plan and the RPA Plan:
Status | 概要 | |||
Pending Plan Upload | This status displays when a case has been generated. | |||
Processing | The plan comparison is processing. | |||
Pending Plan Upload–RETRY | Plan comparison is not possible. The user should review the uploaded files and retry. Possible reasons for this status include uploading the incorrect files. | |||
Comparison Ready–Pass | The Plan Comparison Report is ready for review. All plan comparisons passed the criteria. Some comparisons may be flagged–the user should review the report. | |||
Comparison Ready–Fail | The Plan Comparison Report is ready for review. Some comparisons failed the set criteria–the user should carefully review the report and determine the cause. |
Table 5: Patient status categories for the plan comparison dashboard.
Figure 5: An example of the automatic plan comparison report. Please click here to view a larger version of this figure.
The purpose of the main dashboard (Figure 1) is to provide a quick overview of the status of current patients in the RPA, allow a quick review of completed plans, download completed plans for review and editing in the user's treatment planning system, and provide advanced patient navigation and sorting tools. For patients to appear here, they must have at least one of the following: (1) An Accepted CT scan or (2) an Accepted Service Form. The status categories of current patients can be viewed on the main dashboard (Table 4).
An example lateral field from a whole-brain radiotherapy plan is shown in Figure 6. An example lateral field from a bony landmark-based 4-field box plan for cervical cancer is shown in Figure 7. In both cases, the final plan must be downloaded and then imported into the user's treatment planning system where the results must be reviewed, edited, and recalculated. The RPA also creates a final Plan Report (PDF) that includes the Service Request (see the example in Figure 3), CT approval report, and other details of the treatment plan.
The purpose of the Service Request dashboard (Table 2) is to provide a quick status overview of the Service Request for current patients in the RPA, create a new Service Request, accept a completed Service Request, and edit a Service Request. Patients who have a submitted or Accepted Service Request are displayed on this dashboard, which is accessible to all users of the RPA. However, only users registered in the RPA system as radiation oncologists can Accept a Service Request.
The purpose of the CT dashboard (Table 3) is to provide a quick status overview of CT scans for current patients in the RPA, upload new CT scans, review and Accept CT scans, and add Reference Points to CT scans. Reference points are added to guide the RPA in some specific situations, such as when the user wants to use a non-standard superior border for simple 4-field box plans for cervical cancer. Patients for whom a CT scan has been uploaded are shown here. Any user can view the CT dashboard, but only users registered as Clinical Users can Accept the CT scans.
Once the user has confirmed their final plan, they can export it from their TPS and upload it to the RPA. The purpose of this process is to provide a way to check that the data were properly communicated between different devices.
Patients for whom an RPA Plan has been generated will appear in the plan comparison dashboard. The plan comparison dashboard provides the status categories shown in Table 5 for current patients.
Figure 6: Typical lateral field for whole-brain radiotherapy. This view shows the projections of the structure contours as well as the positions of the main collimators (yellow) and the multi-leaf collimators (blue). Please click here to view a larger version of this figure.
Figure 7: An example lateral field from a bony landmark-based 4-field box plan for cervical cancer. This view shows the positions of the main collimators (yellow) and the multi-leaf collimators(blue). Please click here to view a larger version of this figure.
This protocol outlines the steps in creating automated treatment plans using the RPA. The key steps are (1) CT upload and approval, (2) Service Request completion and approval, (3) plan download and importation into the user's TPS and recalculation of the dose and plan editing, and (4) upload of the final edited plan for comparison with the RPA Plan. The order of the CT approval and Service Request approval are interchangeable. Some plans, specifically volume-modulated arc therapy plans for head and neck and cervical cancer, are created in a two-step process in which additional user interactions and the contours and plans are generated separately. Overall, however, the processes are similar, and we expect that these advanced treatment planning approaches can be changed to a one-step process in the future. The overall clinical acceptability that can be expected from these tools, and from tools that are under development for future versions can be found in our published work10,12,14,15,16,17,18,19,20,21,22.
These tools have several limitations, as described in our earlier work, which investigated risk when deploying the Radiation Planning Assistant in radiotherapy clinics9,23. Although the user interface has been designed to minimize the risk of inappropriate data entry, such as CT images that do not have sufficient field-of-view or errors in manual data entry, there is still potential for error. Specifically, human error, automation bias (over-reliance on the results), and software error are of concern9. Careful review and, if necessary, editing of the automatically generated contours and plans, is essential for the safe use of the Radiation Planning Assistant. In general, these reviews should follow the same process as is followed for the review of clinical plans by physicists and radiation oncologists, although this may be supported by the use of checklists developed specifically to supplement manual review of automatically created treatment plans24.
There are situations where the RPA will not be able to generate a plan and will report an error to the user. In almost all cases, this will be caused by the RPA encountering unexpected data that it cannot interpret, such as insufficient field-of-view or patient positioning (e.g., if a CT image was taken using a supine protocol, but with the patient in a prone position). The user may be able to identify the problem based on where the error is reported. In most cases, these situations can only be remedied by manual contouring or planning. The RPA Team may also be able to review log files to identify the problem.
The RPA has been designed and developed specifically to bring high-quality autocontouring and autoplanning tools to clinics with limited resources, especially those in low- and middle-income countries. We are currently working through the regulatory, legal, and administrative processes that will lead to the RPA being used clinically. Once that happens, we expect to carefully monitor the use and make changes to the workflow or user interfaces in response to any unanticipated risks or other user feedback. The goal is to provide tools that support radiotherapy services, such that local clinical teams can scale their efforts, improving access to high-quality, consistent radiotherapy plans. We hope that this will lead to improved patient outcomes, as well as reduced wait times. Although the current portfolio is limited to cancers of the head and neck, breast, and cervix, as well as whole brain irradiation for brain metastases, we are working on additional treatments that will be incorporated in future versions17,18,19.
The authors have nothing to disclose.
This work was funded by the National Cancer Institute and the Wellcome Trust, with additional support from Varian Medical Systems. Our current system uses Eclipse for treatment planning functions. We would also like to thank Ann Sutton of Editing Services, Research Medical Library, UT MD Anderson Cancer Center. In addition to the institutional funding for RPA development, our research teams receive funding from the Cancer Prevention and Research Institute of Texas (CPRIT) and the Fund for Innovation in Cancer Informatics, The University of Texas MD Anderson Cancer Center.