Session Item

Recently deep neural networks have shown promising results in correcting image quality deficiencies of CBCT images in multiple anatomical locations. Corrected CBCT images, also referred to as synthetic CTs (sCT), can enable daily adaptive proton therapy workflows, which rely on a recalculation of treatment plans on an up-to-date patient anatomy. Compared to 3D-CBCTs, 4D-CBCTs allow to assess the patients breathing and tumour motion on a daily basis but suffer from additional artifacts due to the reconstruction with a very limited projection number. In this study we focused on extending previous deep learning approaches to generate 4D-synthetic CTs using 4D-CBCTs reconstructed from projections acquired with a 3D-protocol. The suitability of 4D-sCTs for proton dose calculations in lung cancer patients was evaluated in terms of image quality and dosimetric accuracy.

CBCT projections and 4D-CTs of 50 lung cancer patients, treated with proton therapy, were used to reconstruct 4D-CBCTs and train a U-net like convolutional neural network to generate 4D-sCTs. Projection phase binning was performed using the Amsterdam Shroud method, and the binned projections were reconstructed into six breathing phases using the MA-ROOSTER reconstruction algorithm. A same-day 4D-CT (ten phases) was used as ground truth image for training and image-quality/dosimetric evaluation. Training of the neural network was performed with deformably registered pairs of 0% phase images of 4D-CBCT and 4D-CT. The dataset was split into a training (25 patients), validation (5) and test set (20). In the test set, image quality (mean absolute error, MAE and mean error, ME) and dosimetric accuracy (3%/3mm gamma analysis) were evaluated on the 0% (inhale) and 50% (exhale) phases. The proton dose distributions were generated by recalculating clinical 2-3 field IMPT treatment plans. Furthermore, we also compared MAE, ME and gamma results of 4D-sCTs to 3D-sCTs from a previous study with thorax patients.

Figure 1 shows an overview of the 0% phase images of the 4D-CBCT, 4D-CT, and the 4D-sCT. An average MAE of 51.5 ± 5.6 HU and ME of 0.2 ± 5.2 HU were observed between 0% phase of 4D-sCT and 4D-CT. Evaluation of the 50% phase resulted in similar results (MAE: 52.4 ± 6.0, ME: 3.3 ± 5.3). The recalculation of clinical treatment plans resulted in average gamma pass ratios of 93.2 ± 3.0 % for the 0% phase and 92.5 ± 3.6 %. Figure 2 shows MAE and Gamma results of 0% and 50% phase for each patient. For comparison, 3D-sCTs resulted in an average MAE of 34.7 ± 7.2 HU and a gamma pass ratio of 93.9 ± 4.6 %.

This study indicates the potential suitability of 4D-sCTs for proton dose calculations in adaptive proton therapy of lung cancer patients. Despite the lower image quality of 4D-CBCTs, dosimetric accuracy of 4D-sCTs is comparable to 3D-sCTs. However further studies are required to investigate reliability and clinical usability of 4D-sCTs.