Source-linked AI summary
Fully Automated Treatment Planning for Head and Neck Radiotherapy using a Voxel-Based Dose Prediction and Dose Mimicking Method
Chris McIntosh, Mattea Welch, Andrea McNiven, David A. Jaffray, Thomas G. Purdie
TL;DR
The paper addresses laborious, subjective radiotherapy planning and the limitations of spatially insensitive dose-volume objectives. It uses atlas-based machine learning to predict per-voxel dose, conditional-random-field inference, and dose mimicking to generate deliverable plans. In 12 right-sided oropharynx patients, automated plans achieved more evaluation criteria in seven patients, equal numbers in four, and improved overall target and organ-at-risk scores.
Problem
Radiotherapy planning is time-consuming and variable, while dose-volume objectives are spatially insensitive and require manual trade-offs for complex head and neck cases.
Method
An atlas-based cARF pipeline predicts probabilistic dose-per-voxel distributions, uses a joint ROI dose prior with a CRF, and applies voxel-based dose mimicking to create deliverable plans.
Results
Automated planning achieved more dose evaluation criteria in 7 patients, equal numbers in 4, and demonstrated higher target coverage and OAR sparing scores than clinical plans.
Takeaways & Limitations
The cARF-CRF[ROI] framework produced automated head and neck plans with more achieved criteria and superior target coverage and OAR sparing compared with clinical plans.
Abstract
from arXiv · showhide
Recent works in automated radiotherapy treatment planning have used machine learning based on historical treatment plans to infer the spatial dose distribution for a novel patient directly from the planning image. We present an atlas-based approach which learns a dose prediction model for each patient (atlas) in a training database, and then learns to match novel patients to the most relevant atlases. The method creates a spatial dose objective, which specifies the desired dose-per-voxel, and therefore replaces any requirement for specifying dose-volume objectives for conveying the goals of treatment planning. A probabilistic dose distribution is inferred from the most relevant atlases, and is scalarized using a conditional random field to determine the most likely spatial distribution of dose to yield a specific dose prior (histogram) for relevant regions of interest. Voxel-based dose mimicking then converts the predicted dose distribution to a deliverable treatment plan dose distribution. In this study, we investigated automated planning for right-sided oropharaynx head and neck patients treated with IMRT and VMAT. We compare four versions of our dose prediction pipeline using a database of 54 training and 12 independent testing patients. Our preliminary results are promising; automated planning achieved a higher number of dose evaluation criteria in 7 patients and an equal number in 4 patients compared with clinical. Overall, the relative number of criteria achieved was higher for automated planning versus clinical (17 vs 8) and automated planning demonstrated increased sparing for organs at risk (52 vs 44) and better target coverage/uniformity (41 vs 31).
1. Introduction
Radiotherapy planning is labor-intensive, iterative, and variable, with head and neck cases posing especially difficult competing target and organ-at-risk priorities. The proposed approach replaces spatially insensitive dose-volume objectives with atlas-based per-voxel dose prediction and dose mimicking.
- Planning challenges: Radiotherapy planning can take hours to multiple days and may produce inter- and intra-institutional variation in practice and quality.Time constraints can also result in sub-optimal plans being delivered, motivating automation for high-throughput planning.
- Planning challenges: Dose-volume objectives are a significant manual step because planners must make subjective trade-offs among competing priorities.Most automated methods infer such objectives from previously treated patients using a limited set of patient and anatomical features.
- Planning challenges: Dose-volume objectives are spatially insensitive within each region of interest, so spatial shaping often requires ad-hoc planning regions and additional objectives.These extra planning regions guide dose spatially but are not evaluated during quality assurance.
- Head and neck complexity: Head and neck planning is challenging because nearby or intersecting organs at risk compete with multiple targets prescribed to up to three dose levels.The site has motivated efforts to improve planning standardization and efficiency.
- Proposed approach: The proposed atlas-based method infers dose-per-voxel for a novel planning image from patient geometry, ROI features, and global and local image-intensity features.Its cARF framework produces probabilistic voxel-level dose estimates, creating a spatial dose objective rather than a region-level estimate.
- Proposed approach: The pipeline combines probabilistic spatial dose prediction, conditional-random-field scalarization under a dose prior, and dose mimicking to produce deliverable plans without specified dose-volume objectives.The study applies this fully automated framework to right-sided oropharynx head and neck radiotherapy.
2. Methods and Materials
The study trained and tested atlas-based voxel-dose prediction on right-sided oropharynx patients using image and ROI features, atlas selection, conditional-random-field inference, and voxel-based dose mimicking. The resulting predicted distributions were converted into deliverable plans using collapsed cone convolution dose calculations.
- Patients and plans: The dataset contained 66 right-sided oropharynx patients, divided into 54 training image-plan pairs and 12 randomly selected independent testing instances.The patients were planned and treated between 2011 and 2013 according to the institutional protocol.
- Patients and plans: Clinical plans used Pinnacle3 planning and Varian TrueBeam delivery, comprising 59 IMRT plans and 7 VMAT plans.IMRT used nine step-and-shoot beams, while VMAT used two 330-degree arcs with 3-degree gantry spacing.
- Patients and plans: Patients were treated in 35 fractions with prescriptions of 7000 cGy to high-risk and 5600 cGy to elective target volumes.Seven patients also received 6300 cGy to an intermediate-risk target, and five had bilateral high-risk target volumes.
- Automated spatial dose prediction pipeline: The prediction pipeline characterizes patient geometry, ROI shape, and global and local image appearance with gradient-based and texture features, then uses regression and density estimation for atlas selection.The learned models estimate dose-feature relationships voxel by voxel and select relevant atlases for novel images.
- Automated spatial dose prediction pipeline: The model learns a joint dose distribution prior over all ROIs and their intersections, unlike a typical DVH prior defined cumulatively per ROI.A conditional random field selects the most likely spatial dose-per-voxel assignment while adhering to the inferred prior.
- Dose mimicking: Voxel-based dose mimicking converted predicted and clinical dose distributions into deliverable treatment plans through iterative collapsed cone convolution calculations.The clinical beam arrangement was retained for mimicking, using nine beams for IMRT and two partial arcs for VMAT.
- Dose mimicking: Dose mimicking matched dose at each voxel across the dose grid rather than optimizing dose distributions per ROI or using dose-evaluation objectives.The same spatial objectives and weights were used for all patients and methods.
3. Results
Across 54 training and 12 independent testing patients, four automated pipeline variants were evaluated using dose-criteria compliance, comparative dose outcomes, runtime, and feature importance. cARF-CRF[ROI] performed best overall, with dose mimicking improving its achieved criteria.
- Dose evaluation criteria: cARF-CRF[ROI] achieved 140 dose evaluation criteria overall, compared with 136 for cARF-CRF, 132 for cARF[ROI], and 131 for clinical.Dose mimicking increased cARF-CRF[ROI] from 135 to 140 achieved criteria and clinical from 127 to 131.
- Dose evaluation criteria: 17 versus 8 dose evaluation criteria were uniquely achieved by cARF-CRF[ROI] compared with clinical, while the comparison with cARF-CRF was 9 versus 5.At the patient level, cARF-CRF[ROI] achieved more criteria than clinical in 7 patients, fewer in 1, and the same number in 4.
- Target coverage and OAR sparing: 93 versus 75 dose evaluation criteria were better achieved by cARF-CRF[ROI] than clinical, compared with 90 versus 78 against cARF-CRF.The comparison counted lower doses for sparing criteria and higher doses for target coverage criteria.
- Computational performance: Approximately 10 minutes per patient was required for the average cARF-ROI dose inference pipeline on the reported compute server.Runtime varied with CT slices, dose-grid size, and the number of ROI voxels.
- Feature importance: Feature importance was assessed for ARFs with and without OAR features using out-of-bag permutation accuracy during training.Without OAR features, the algorithm relied more on larger-scale gradient and additional target features.
4. Discussion and Related Work
The cARF-CRF[ROI] pipeline produced deliverable dose-mimicked plans that often matched or improved clinical plans, while results exposed sensitivity to atlas selection, ROI inputs, and training-data scope.
- 111 cGy ± 184 cGy was the mean absolute dosimetric discrepancy between pre- and post-mimicked clinical dose distributions.
- cARF-CRF[ROI] achieved more dose evaluation criteria in 7 patients, the same number in 4, and fewer in 1 than clinical plans.
- 52 versus 44 target coverage criteria and 41 versus 31 OAR-sparing criteria favored mimicked cARF-CRF[ROI] plans over clinical plans.
- cARF-CRF[ROI] was the only proposed cARF method to achieve more dose evaluation criteria after dose mimicking, suggesting more realistic predicted distributions.It combines OAR features for atlas selection with a conditional random field for spatial dose distribution.
- Without OAR ROI features, cARF-CRF selected better atlases but mapped their dose distributions less accurately into achievable plans; after mimicking, cARF-CRF[ROI] performed better in 6 patients.The authors associate this reversal with possible over-fitting to atlas-selection data and contouring variation.
- The method’s scope is constrained by limited ROI inputs, a training database requirement, and evaluation using clinical beam orientations.The study used 54 training patients; varying beam orientations across treatment sites remain a future consideration.
5. Conclusions
The cARF-CRF[ROI] method evaluates achievable per-voxel dose prediction within a fully automated planning framework, generating complete treatment plans without dose-volume objectives. In 12 head and neck patients, it produced superior target coverage and OAR sparing compared with clinical plans.
- The cARF methods require images, target delineations, and a limited number of anatomical OARs, while generating spatial dose objectives instead of dose-volume objectives.This replaces a manual step in the canonical planning process.
- The cARF-CRF[ROI] method achieved more dose evaluation criteria and superior target coverage and OAR sparing compared with clinical plans in 12 head and neck patients.
- The cARF-CRF[ROI] method predicted spatially accurate dose distributions that were readily mimicked into complete treatment plans.
- Feature-based machine learning and atlas selection make the method highly generalizable to other treatment sites and treatment modalities.