Session Item

Atlas based machine learning (ML) for radiation treatment planning effectively creates new treatment plans by selecting the most anatomically similar patients (i.e., atlases) from a training database and mapping their dose distributions to a novel patient’s unique anatomy. We investigated whether an atlas-based ML model, trained only on computed tomography (CT) imaging, could generate clinically applicable treatment plans in the MR guided setting by directly predicting dose on MR imaging without model retraining.

We included contoured CT and 3D T2 MR imaging from prostate cancer patients (n=55) treated on a 1.5T MR-linac. For each patient, we generated moderately hypofractionated (60 Gy in 20 fractions) VMAT treatment plans on both images using an atlas-based ML treatment planning model, trained on CT imaging only. Statistically significant differences between dose distributions of MR- and CT-based treatment plans were identified using DVH metrics, as per institutional evaluation criteria, and Wilcoxon Signed Rank tests (α = 0.05). To determine if these dosimetric differences were due to anatomical differences between images, they were compared to differences in relative PTV overlap of the rectum and bladder walls between images, calculated as the volume of OAR-PTV intersection divided by the OAR volume. To determine if changing the input image influenced model predictions (i.e., by selecting dosimetrically differing atlases in anatomically similar cases), differences in overlap and DVH metrics between patients and their selected atlases were compared for each image type.

Statistically significant changes in dose-volume metrics between MR- and CT-based plans were identified for PTV D99, bladder wall D30 and D50, and rectum wall D50 (Figure 1). Differences in the amount of PTV overlap between the MR and CT images moderately correlated (r > 0.7, p < 0.001) with differences in dose-volume metrics (Figure 2a). Differences in overlap between a patient and their selected atlases correlated (r > 0.7, p < 0.001) with the dose difference between them (Figure 2b), and this relationship was consistent between CT- and MR- based plans

Figure 1: Distribution of dose-volume metrics for MR- and CT-based treatment plans. Dotted lines indicate institutional evaluation criteria. Asterisks indicate statistically significant changes in dose.

Figure 2: The relationship between dosimetric and PTV overlap differences for a) MR- and CT- based ML plans, and b) each treatment plan and its respective selected atlases.

The dosimetric differences observed between MR- and CT-based treatment plans are attributable to anatomical differences between the images, rather than altered ML behaviour resulting from changing the input imaging modality to MR. Therefore, we have demonstrated that an atlas-based ML model, previously trained on CT imaging only, can be safely repurposed for use in the MR guided setting.