Cone-Beam CT to CT Image Translation Using a Transformer-Based Deep Learning Model for Prostate Cancer Adaptive...

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Cone-beam computed tomography (CBCT) offers real-time imaging for image-guided radiation therapy but suffers from lower image quality compared to planning CT (pCT).
• This quality deficit limits the application of CBCT in adaptive radiotherapy (ART), which requires high-fidelity images for treatment adjustments.
• Enhancing CBCT image quality is crucial for improving the precision and effectiveness of ART.

Purpose of the Study:

• To enhance CBCT image quality for radiation therapy using a transformer-based deep learning model, SwinUNETR.
• To compare the performance of SwinUNETR against a conventional convolutional neural network (CNN) model, U-net, for CBCT-to-CT image synthesis.
• To evaluate the dosimetric impact of the synthesized CT images for photon therapy.

Main Methods:

• A retrospective study of 260 prostate radiotherapy patients was conducted, with data split for training (245 patients) and independent testing (15 patients).
• A CycleGAN framework was utilized to generate synthetic CT (sCT) images from CBCT, with SwinUNETR and U-net serving as generators.
• Image quality was assessed using mean absolute error of CT numbers, and dosimetric accuracy was evaluated via gamma analysis and dose-volume histogram (DVH) comparisons against pCT.

Main Results:

• SwinUNETR achieved a lower mean absolute error (64.69 HU) compared to U-net (73.49 HU) and raw CBCT (84.07 HU) when compared to pCT.
• Gamma analysis demonstrated superior dose agreement between pCT and SwinUNETR-derived sCT plans than with U-net derived plans.
• DVH parameters calculated on SwinUNETR-based sCT images showed less than 1% deviation from pCT plans.

Conclusions:

• The transformer-based SwinUNETR model significantly outperforms the CNN-based U-net model in generating high-quality synthetic CT images from CBCT.
• Improved synthetic CT image quality facilitates more accurate dose calculations, crucial for adaptive radiotherapy.
• This study highlights the potential of transformer architectures for advancing CBCT-to-CT image translation in radiation oncology.