Session Item

To demonstrate that fully automated volumetric modulated arc therapy (VMAT) dose distributions for oropharyngeal cancer patients can be generated with machine learning optimization (MLO) planning, with similar quality as the clinical ‘dosimetrists-optimized’ dose distributions, further indicated as reference plans.

MLO planning involved training of a model using 60 oropharyngeal cancer patients, which was used to predict the voxel dose for new patients. CT scans, structures and dose distributions of 99 consecutive primary oropharyngeal cancer patients, previously treated with dual arc VMAT, were retrieved from our clinical database.  Image and contour features were extracted and atlas regression forests (ARF), prediction random forests (pRF) and conditional random fields (CRF) were trained. Using the trained model, spatial dose distributions were predicted and optimized to generate clinical treatment plans while adhering to the predicted dose. Validation was performed with 39 oropharyngeal cancer patients to tune model settings using both target and organ at risk (OAR) quality measures. Clinical machine learning plans and reference plans were compared by means of adequate target coverage (D₉₈≥95%), dose on OARs (D_0.1, D_mean), normal tissue complication probability (NTCP) values (xerostomia, dysphagia and tube feeding dependence) and planning time. Two-tailed p-values were calculated by a paired Wilcoxon signed–rank test and a Bonferroni correction (α=0.05).

The predicted dose was in agreement with the reference dose for all plans, see table 1. Validation showed that it was possible to incorporate clinical requirements in the model settings. In the final settings, both the clinical machine learning and reference plans had adequate target coverage in 37/39 (95%) and acceptable maximum OARs dose in 38/39 (97%) of the plans. The average sum NTCP was 84.5% (±17.8) and 84.9% (±19.4) for clinical machine learning plans and reference plans, respectively. Planning time for reference plans took around 240 minutes and clinical machine learning plans were generated in 65 minutes (±11.2), with negligible hands-on time. Figure 1 shows the average dose volume histogram (DVH) of all reference and clinical machine learning plans.

In this study, we demonstrate that clinical machine learning plans for oropharyngeal cancer patients have comparable plan quality to the reference plans, while more consistent and much more efficient to generate. In future work, we will perform a blinded prospective clinical study on MLO planning versus manual planning. MLO planning can be also be used for automated generation of multiple treatment plans, for both photons and protons.