Purpose: Identifying head neck cancer (HNC) patients who would benefit most from on-line adaptive radiotherapy (ART) is of great importance for clinical workflow management due to tremendous resources required for repeated contouring/replanning/quality assurance and limited benefit for a sub-group of patients with minimal anatomic change. In this study, we built a vision-transformer (ViT) based deep neural network to predict radiotherapy induced anatomic change of HNC patients for assisting in early identification of HNC patients who can benefit from on-line ART.
Methods: We retrospectively collected data from 260 HNC patients (195 for training, 65 for validation) treated with definitive RT/CRT at our institution. These data comprise planning CTs, dose, and CBCTs acquired at 1 week (CBCT1) and 3 weeks (CBCT3) of starting RT. After image registration, resampling, and normalization, we cropped the images to regions including oral cavity, pharynx, and cervical esophagus, and used them as the model inputs and targets. An UNet-style ViT network was used to learn the spatial correspondence and contextual information from embedded image patches of CT, dose, and CBCT1 images. The deformation vector field between CBCT1 and CBCT3 was then estimated by the model as the prediction of anatomic change, and deformed CBCT1 was used as the prediction of CBCT3. We also generated binary human body masks in head neck region based on CT and CBCTs for volumetric change evaluation.
Results: Image and volumetric similarity metrics including mean square error, structural similarity index, dice coefficient, average surface distance, and net volumetric change were used to measure the similarity between the predicted CBCT3 and real CBCT3. The proposed method yielded better similarity than CT/CBCT1 to CBCT3 for all the metrics.
Conclusion: The proposed method showed promising performance for predicting radiotherapy induced anatomic change, which has the protentional for early identification of HNC patients who benefit from ART.
Not Applicable / None Entered.
IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)