Samenvatting
Purpose/Objective
In proton therapy it is common practice to acquire weekly verification CTs to monitor treatment progress and recalculate treatment plans on updated patient anatomy. For daily adaptive proton therapy workflows however, repurposing in-room images such as cone-beam CTs (CBCT), is more suitable since it is not adding to the clinical workload and does not cause any additional dose burden to the patient. CBCT images, routinely acquired for pre-treatment position verification, provide a daily representation of the patient anatomy but suffer from severe imaging artefacts preventing accurate dose calculations. Recently deep neural networks have shown promising results to correct CBCT images and generate high quality synthetic CTs (sCT), for proton dose calculations.
Therefore, the aim of this study was to compare weekly rCT and daily sCT images of head and neck cancer patients to investigate the dosimetric accuracy of CBCT-based sCTs generated by a neural network.
Materials and Methods
A dataset of 30 head and neck cancer patients was utilized to generate synthetic CTs from daily pre-treatment patient alignment CBCTs using a previously developed and trained UNet deep convolutional neural network. Afterwards, clinically used proton treatment plans were recalculated on sCTs and weekly rCTs to evaluate the dosimetric accuracy of sCTs. Dose to clinical target volumes (CTV) and selected organs-at-risk (OAR) were compared between pCTs and both weekly rCTs and same-day sCTs by calculating mean dose differences. The investigated organs-at-risk include submandibular glands, pharyngeal constrictor muscles, parotid glands and the oral cavity.
Results
Figure 1 shows the mean relative dose differences between sCT/pCT and rCT/pCT pairs per ROI. The best agreement between pCT and rCT/sCT was observed for the low dose CTV (CTV 5425) with mean dose difference values of 0.3±0.2 % [0.18±0.15 Gy](rCT) versus 0.5±0.7 % [0.32±0.45 Gy](sCT), and for 0.2±0.2 % [0.13±0.11 Gy](rCT) versus 0.3±0.2 % [0.19±0.16 Gy](sCT) for the high dose CTV (CTV 7000). For all OARs, significantly larger dose differences were found than for the CTVs. The largest difference between sCT and pCT doses were observed in the left parotid gland with 14.7±16.4 % [1.61±1.31 Gy], compared to 8.4±9.2 % [1.01±0.87 Gy] between rCT and pCT, and in the left submandibular gland with 4.1±5.3 % [1.32±1.37 Gy](rCT) compared to 8.3±11.6 % [2.43±2.69 Gy](sCT). Overall, rCTs showed lower dose differences for all regions of interest. Figure 2 shows a comparison of daily sCT/pCT and weekly rCT/pCT dose differences for target volumes and OARs of the entire treatment of an exemplary patient.
Conclusion
The deep learning based sCTs showed high agreement for target volume doses (
In proton therapy it is common practice to acquire weekly verification CTs to monitor treatment progress and recalculate treatment plans on updated patient anatomy. For daily adaptive proton therapy workflows however, repurposing in-room images such as cone-beam CTs (CBCT), is more suitable since it is not adding to the clinical workload and does not cause any additional dose burden to the patient. CBCT images, routinely acquired for pre-treatment position verification, provide a daily representation of the patient anatomy but suffer from severe imaging artefacts preventing accurate dose calculations. Recently deep neural networks have shown promising results to correct CBCT images and generate high quality synthetic CTs (sCT), for proton dose calculations.
Therefore, the aim of this study was to compare weekly rCT and daily sCT images of head and neck cancer patients to investigate the dosimetric accuracy of CBCT-based sCTs generated by a neural network.
Materials and Methods
A dataset of 30 head and neck cancer patients was utilized to generate synthetic CTs from daily pre-treatment patient alignment CBCTs using a previously developed and trained UNet deep convolutional neural network. Afterwards, clinically used proton treatment plans were recalculated on sCTs and weekly rCTs to evaluate the dosimetric accuracy of sCTs. Dose to clinical target volumes (CTV) and selected organs-at-risk (OAR) were compared between pCTs and both weekly rCTs and same-day sCTs by calculating mean dose differences. The investigated organs-at-risk include submandibular glands, pharyngeal constrictor muscles, parotid glands and the oral cavity.
Results
Figure 1 shows the mean relative dose differences between sCT/pCT and rCT/pCT pairs per ROI. The best agreement between pCT and rCT/sCT was observed for the low dose CTV (CTV 5425) with mean dose difference values of 0.3±0.2 % [0.18±0.15 Gy](rCT) versus 0.5±0.7 % [0.32±0.45 Gy](sCT), and for 0.2±0.2 % [0.13±0.11 Gy](rCT) versus 0.3±0.2 % [0.19±0.16 Gy](sCT) for the high dose CTV (CTV 7000). For all OARs, significantly larger dose differences were found than for the CTVs. The largest difference between sCT and pCT doses were observed in the left parotid gland with 14.7±16.4 % [1.61±1.31 Gy], compared to 8.4±9.2 % [1.01±0.87 Gy] between rCT and pCT, and in the left submandibular gland with 4.1±5.3 % [1.32±1.37 Gy](rCT) compared to 8.3±11.6 % [2.43±2.69 Gy](sCT). Overall, rCTs showed lower dose differences for all regions of interest. Figure 2 shows a comparison of daily sCT/pCT and weekly rCT/pCT dose differences for target volumes and OARs of the entire treatment of an exemplary patient.
Conclusion
The deep learning based sCTs showed high agreement for target volume doses (
| Originele taal-2 | English |
|---|---|
| Status | Published - 2022 |
| Evenement | ESTRO 2022 - Copenhagen, Denmark Duur: 6-mei-2022 → 10-mei-2022 |
Conference
| Conference | ESTRO 2022 |
|---|---|
| Land/Regio | Denmark |
| Stad | Copenhagen |
| Periode | 06/05/2022 → 10/05/2022 |