Voxel-Based Dose Prediction with Multi-Patient Atlas Selection for Automated Radiotherapy Treatment Planning

Chris McIntosh,Thomas G. Purdie

DOI: https://doi.org/10.1088/1361-6560/62/2/415

2016-08-16

Abstract:Automating the radiotherapy treatment planning process is a technically challenging problem. The majority of automated approaches have focused on customizing and inferring dose volume objectives to used in plan optimization. In this work we outline a multi-patient atlas-based dose prediction approach that learns to predict the dose-per-voxel for a novel patient directly from the computed tomography (CT) planning scan without the requirement of specifying any objectives. Our method learns to automatically select the most effective atlases for a novel patient, and then map the dose from those atlases onto the novel patient. We extend our previous work to include a conditional random field for the optimization of a joint distribution prior that matches the complementary goals of an accurately spatially distributed dose distribution while still adhering to the desired dose volume histograms. The resulting distribution can then be used for inverse-planning with a new spatial dose objective, or to create typical dose volume objectives for the canonical optimization pipeline. We investigated six treatment sites (633 patients for training and 113 patients for testing) and evaluated the mean absolute difference (MAD) in all DVHs for the clinical and predicted dose distribution. The results on average are favorable in comparison to our previous approach (1.91 vs 2.57). Comparing our method with and without atlas-selection further validates that atlas-selection improved dose prediction on average in Whole Breast (0.64 vs 1.59), Prostate (2.13 vs 4.07), and Rectum (1.46 vs 3.29) while it is less important in Breast Cavity (0.79 vs 0.92) and Lung (1.33 vs 1.27) for which there is high conformity and minimal dose shaping. In CNS Brain, atlas-selection has the potential to be impactful (3.65 vs 5.09), but selecting the ideal atlas is the most challenging.

Medical Physics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the process of automating dose prediction in radiotherapy planning. Specifically, the authors propose a voxel - dose prediction method based on multi - patient atlas selection, aiming to directly predict the dose distribution of each voxel from the patient's CT scan images without specifying any dose - volume objectives. This method achieves dose prediction by automatically selecting the most effective atlases and mapping the doses in these atlases to new patients. In addition, the authors introduce a conditional random field (CRF) model to optimize the joint probability distribution prior to match the complementary goals between the accuracy of the spatial dose distribution and the required dose - volume histogram (DVH). The goal of this method is to provide a solution that can accurately predict the dose distribution without manually setting targets, which is applicable to multiple clinical sites and treatment modalities and can be achieved using a single framework.

### Main contributions of the paper
1. **Extended previous methods**: Combined features based on spatial dose distribution and features derived from DVH, as well as joint dose priors, to improve the dose prediction method.
2. **Validated new clinical sites**: Validated on three new clinical sites (central nervous system - brain, lung, and rectum).
3. **Introduced the conditional random field model**: Transformed the probability dose distribution into a scalar dose distribution to make it conform to the predicted joint probability function prior (similar to DVH), thereby generating a spatially appropriate dose distribution while ensuring that the dose is appropriately distributed within the demarcated ROI.

### Method overview
- **Feature extraction**: Extract texture features at multiple scales, rotation - invariant features, and target - specific features from CT images.
- **Atlas - to - image mapping**: Use the Attribute Regression Forest (ARF) to learn the relationship between each training image and dose pair while calculating the probability density function estimate.
- **Atlas selection**: Predict the dose prediction accuracy of each ARF on other training images through cross - validation and use density estimation to learn to predict the observed accuracy. Finally, select the k ARFs with the highest prediction accuracy for dose prediction of new images.
- **Conditional random field dose estimation**: Use the CRF model to infer the most likely spatial dose distribution from the probability dose distribution to make it conform to the joint dose prior.

### Results
The paper conducted experiments on six different treatment sites (breast cavity, whole breast, central nervous system - brain, prostate, lung, and rectum). The results show that this method performs better than previous methods in most sites, especially in sites that require precise dose distribution (such as whole breast, prostate, and rectum).