Session Item

The ability to calculate dose on CBCTs, using synthetic CTs (sCTs), has the potential to make the adaptive radiotherapy (ART) pathway more efficient and remove subjectivity from the process. Implementing sCTs generated from CBCTs into the ART pathway would also reduce CT scanner workload and allow adaptive treatment plans to be delivered more quickly.

This study assessed the dosimetric and Hounsfield units (HU) similarity of CBCT-based sCTs compared to CTs as well as the sCT generation time. sCTs were generated using a commercially available treatment planning system.

Fifteen head and neck rescan patients were used to assess four methods of sCT generation using RayStation Research version 9B. Each patient’s planning CT (pCT), rescan CT (rCT), and the first CBCT after the rCT were obtained, using the rCT as the comparator. The CBCT was deformed to the rCT geometry (dCBCT) and used as the input for sCT generation.

Method 1 deformably registered the pCT to the dCBCT. Method 2 assigned the range of dCBCT intensity values to six mass density values. Method 3 iteratively removed low-frequency artefacts and assigned a HU function to the dCBCT values. Method 4 used a cycle generative adversarial network (cycleGAN) machine learning model (independently trained using 45 head and neck patient dCBCTs and pCTs) to generate an sCT. Methods 1, 3, and 4 are currently RayStation Research only scripted methods.

A treatment plan conforming to the local clinical protocol was created on each rCT and recalculated on each sCT. Planning target volume (PTV) and organ at risk (OAR) structures were contoured by clinicians on the rCT to allow assessment of dose-volume histogram (DVH) statistics. The mean absolute error (MAE) of the HU, dose differences of PTV and OAR structures (high-dose PTV, low-dose PTV, spinal canal, larynx, brainstem, and parotids) at clinically relevant DVH points, and global gamma index analysis (2%/2 mm) were used to assess the differences between the sCT and rCT. sCT generation time, including validation, was also recorded.

For methods 1, 2, 3, and 4 the MAE, gamma index analysis, and generation time were: 59.7 HU, 100.0%, and 143 s; 164.2 HU, 95.2%, and 232 s; 75.7 HU, 99.9%, and 153 s; and 79.4 HU, 99.8%, and 112 s respectively. All assessed dose differences were <0.3 Gy except for method 2 (<0.5 Gy). An example of the dose differences between the rCT and sCTs are shown in Figure 1.

Figure 1 - a) Dose distribution (Gy) on the rCT. b), c), d), and e) Dose differences (% of prescription dose) vs. a) for methods 1, 2, 3, and 4 respectively. Positive values indicate an underdose relative to a).

All methods were considered clinically viable. Method 4, the machine learning method, was found to be most suitable for clinical implementation due to its high dosimetric accuracy and short generation time.

Further investigation is required to assess these methods in situations where the CBCT and CT are significantly different and for other anatomical sites.