Session Item

Current research in radiotherapy treatment planning is directed toward 3D dose predictions. Usually a dose mimicking optimization process is used to translate the 3D dose predictions into treatment device parameters, which is time consuming and based on 1D DVH information. Recently, fluence predictions based on 3D dose information using CNNs have been proposed as an alternative. Proof of principle publications create a model for one specific class solution (beam setup and treatment device). Here, we investigate the possibility to transfer a CNN-based fluence prediction model for lung IMRT trained on our Hacyon class solution to a TrueBeam STx class solution with another beam setup, for which no training data is available.

A dataset of 153 lung cancer patients, treated with IMRT, is available. For these patients the beam dose distribution and fluence map for each field is exported from Eclipse^TM. The training (n=97) and validation set (n=30) contain patients treated on a Halcyon device with energy 6MV-FFF with either SX1 (1cm leaf width) or SX2 (0.5cm virtual leaf width) and a 9 beam setup, see Figure 1. In the test set (n=26), the patient plans are created for a TrueBeam STx device (0.25 cm leaf width) with the same energy, but a 6 beam setup, see Figure 1. The CNN has 1 input: the 3D beam dose distribution, rotated according to the beam’s axis (gantry and collimator angle) and one output: the fluence map of the corresponding beam. A 2D Unet is used and the 3D dose input is given as multiple 2D channels to the network. After fluence prediction, the fluences are imported in the treatment planning system for dose calculation.

Per patient, the workflow takes on average 137s for preprocessing of the doses and 0.96s for fluence prediction. The results of the clinical plans vs. the plans obtained after fluence prediction can be found in Figure 2 and Table 1. All plans are normalized to have a dose of 1 Gy as mean dose in the PTV. The differences between the mean doses for the clinical constraints between the clinical and predicted plan are mostly smaller than 2% of the prescribed dose. The largest differences can be found at D(0.035cc) of PTVs with mean differences of 4% of the prescribed dose. 

Table 1: Mean±std over the test set, relative to the prescribed dose.

It is possible to transfer a fluence prediction model, trained for a 9 beam Halcyon class solution, to a 6 beam TrueBeam STx class solution, of which insufficient data is available to train a separate network.