Author: Lei Ren, Jie Zhang 👨🔬
Affiliation: University of Maryland School of Medicine 🌍
Purpose: 4D-CBCT is valuable for imaging anatomy affected by respiratory motions to guide radiotherapy delivery. However, 4D-CBCT often has undersampled projections acquired in each respiratory phase due to the limit in scanning time and dose, severely impacting the image quality. This study aims to enhance its image quality by using a deep learning model to predict full-view projections.
Methods: A deep learning model was developed to use two projections at different angles as the inputs and predict projections at other specified angles. Specifically, the model uses a U-net backbone with our self-designed convolutions to fit the rotation-induced signal variation. The projection at the angle of β (uβ) is generated according to its two neighbors (uα1 and uα2) in sparse views (SV). β-α1 and β-α2 are inputted to our model to generate regulators to modify the model weights, which is to adaptively adjust the contribution ratio of uα1 and uα2. The model was trained and validated using the Elekta 4D-CBCT data from real patients provided by the AAPM SPARE challenge. {uα1, uα2} and uβ composed one sample pair. There were 2695 sample pairs from No.1~4 patients for training and 669 pairs from No.5 patient for validation. The generated projections and its reconstructed volumes were compared with the true ones in the challenge data using root mean square error (RMSE), peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM).
Results: The predicted projections show comparable quality to the ground truth with a RMSE of 0.154±0.087, a PSNR of 32.339±4.758 dB and a SSIM of 0.992±0.011. The volumes are enhanced with a RMSE of 0.002±0.001 (SV: 0.033±0.023), a PSNR of 38.911±3.532 dB (SV: 20.926±10.838) and a SSIM of 0.964±0.019 (SV: 0.407±0.384).
Conclusion: The proposed model predicts full-view projections from a sparse-view 4D-CBCT. It enables high-quality imaging with reduced scan time and dose.