Fig. 5.

Dose difference comparison shown as the mean percentage dose error evaluated for multiple dose-volume histogram (DVH) parameters for n = 17 test patients. In general, 2D dose distribution predicted by the baseline U-Net model led to a poor dose coverage of the target volume (a). The densely connected model trained in 2.5D resulted in an overall lower dose difference for the target, body, and most of the organs at risk (OAR) structures, when compared to the 2D model (b). The mean percentage dose error stayed within 2.6%, 1.1% and 1.9% for OAR, body, and target structures respectively, whereas dose distributions predicted by the 2D baseline U-Net stayed within 3.4%, 2.7% and 4.4% for OAR, body, and target. Using our proposed treatment panning workflow, dose predictions derived from the densely connected U-Net could successfully be transformed into deliverable treatment plans, staying within a mean percentage dose error of 4.4%. Dose coverage of the target volume, body dose and the dose in most of the OARs were close to the ground truth (GT) dose distributions (c). The mean percentage dose error between the predicted dose distributions of the 2.5D densely connected U-Net and the deliverable treatment plans was found to be 2.5%, 3.1% and 2.7% for OAR, body, and target volumes, respectively (d).