Automatic adaptive radiotherapy triggering based on CBCT using deep learning for esophageal cancer underwent volumetric modulated arc therapy.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medical Engineering, 1st Affiliated Hospital of Wenzhou Medical University, Wenzhou, China.
- Department of Radiation and Medical Oncology, 1st Affiliated Hospital of Wenzhou Medical University, Wenzhou, China.
- Department of Radiation Oncology, The Affiliated Suzhou Hospital of Nanjing Medical University, Suzhou Municipal Hospital, Suzhou, China.
- School of Basic Medical Science, Wenzhou Medical University, Wenzhou, China.
Abstract
Esophageal cancer (EC) is a highly fatal malignancy for which radiotherapy plays a critical role in treatment. However, anatomical changes during radiotherapy can lead to increased doses to surrounding organs at risk (OARs). Adaptive radiotherapy (ART) is a strategy that incorporates patient-specific anatomic changes into treatment plan modification to minimize overdose to surrounding healthy tissues. Identifying appropriate triggers with in-room verification imaging is critical to maximize the benefits of ART and prevent additional imaging dose to a certain extent for patients without compromising clinical resources or operations. To develop an automatic ART triggering procedure by predicting the patient's anatomical changes and their resulted dose-volume metrics differences on OARs using deep learning (DL) based on cone-beam computed tomography (CBCT) to efficiently and effectively balance the benefit and frequency of ART. A DL network was first trained to automatically segment OARs on the CBCT of 136 EC patients underwent volumetric modulated arc therapy (VMAT). The dose distributions on CBCT with the automatically contoured OARs were predicted with Unet. A set of trigger criteria based on OAR dosimetry deformed form original treatment plan was established to assess replanning necessity. The feasibility and accuracy of automatic segmentation and dose prediction on CBCT were verified with rescan CT (rCT) and CBCT at the same time point. The average dices of the automatic segmentation model for the lung, heart, and spinal cord were 0.92, 0.91, and 0.80, respectively. The predicted dose distributions on CBCT were close to mapped dose distributions. The ART trigger decision agreement between rCT and CBCT was 81.8%. CBCT with automatic segmentation OARs achieved an area under curve, accuracy, sensitivity, and specificity of 0.86, 0.82, 1.0, and 0.71 in the triggering of ART for EC patients, respectively. An automatic ART triggering procedure was established based on CBCT directly for EC patients underwent VMAT. It is a feasible and promising ART methods to improve the management of EC patients without additional patients appoints and resources.