Session Item

The integration of artificial intelligence will play a fundamental role for further developing radiotherapy in the upcoming years. However, it requires the implementation of dedicated quality assurance (QA). This applies in particular in the context of MR-only radiotherapy, where the simulation CT required for dose calculation is substituted by a synthetic CT (sCT). MR-only radiotherapy simplifies the treatment planning workflow, but poses new challenges for patient specific QA (PSQA).

This study retrospectively analysed 20 patients treated at a hybrid MR-Linac for abdominal lesions to assess different PSQA techniques for sCT. The patients were selected to equally cover four specific subgroups presenting the following features along the beam path: (a) standard cases, (b) air pockets, (c) lungs and (d) implants. Each patient underwent an MR simulation at the MR-Linac (sequence TrueFISP, field 0.35T) followed by a CT. The retrospective simulation of an MR-only workflow assumed that the electron density map for dose calculation was computed from a sCT generated using a neural network (CycleGAN) from the MR data. Figure 1 displays the four methods investigated to generate an independent electron density map for PSQA: (i) water override of the body mask, (ii) manual bulk override of five tissue classes, (iii) sCT from an independent neural network (NN) based on pix2pix model and (iv) deformed CT from the CT simulation (dCT). IMRT plans were recalculated with preset MU for the four PSQA approaches and DVH parameters were extracted. The statistical significance was assessed with a non-equivalence test for dependent paired samples (TOST-P).

The time allocation for the PSQA task was less than 2 minutes for (i), up to 20 and 10 minutes for (ii) and (iii), respectively. The times between the MR and CT simulation for (iv) were patient dependent and limited to 30 minutes. The dCT was the only method including the artificial implants in the electron density map. Figure 2 reports the observed differences of the GTV mean dose between the calculation performed on the reference sCT and the PSQA. Deviations within 2% were observed only for (iii) and (iv) across the whole cohort. In absence of large air pockets and lung tissue within the beam path, (i) and (ii) also provided valuable results. Equivalence testing (p=0.05) of multiple DVH parameters confirmed that the independent NN and dCT can provide PSQA applicable to the 1% level for the sub-groups (a) and (b), 1.5% for (c) and 2% for (d).

Four independent strategies for performing PSQA of sCT in the context of MR-only radiotherapy were investigated. The use of an independent NN generating as sCT for dose verification was demonstrated to be applicable for PSQA if testing DVH parameters to the 2% equivalence level. The PSQA task was performed within 10 minutes for each patient, which meets the requirements for the integration within the planning workflow.