MBST-Driven 4D-CBCT reconstruction: Leveraging swin transformer and masking for robust performance
Nannan Cao1, Qilin Li2, Kangkang Sun1
1Department of Radiotherapy, The Affiliated Changzhou NO.2 People's Hospital of Nanjing Medical University, Changzhou, 213003, PR China; Jiangsu Province Engineering Research Center of Medical Physics, Changzhou, 213003, PR China; Center for Medical Physics, Nanjing Medical University, Changzhou, 213003, PR China; Key Laboratory of Medical Physics in Changzhou, Changzhou, 213003, PR China.
A novel Mask-based Swin Transformer network (MBST) significantly enhances 4D cone-beam computed tomography (4D-CBCT) image quality. This deep learning approach improves reconstruction accuracy and detail, even with limited scanning data, offering better diagnostic capabilities.
Area of Science:
- Medical Imaging
- Deep Learning
- Radiotherapy
Background:
- 4D cone-beam computed tomography (4D-CBCT) is crucial for image-guided radiotherapy, but its image quality is often limited by motion artifacts and sparse data acquisition.
- Improving 4D-CBCT reconstruction quality is essential for accurate tumor targeting and dose delivery during radiation therapy, especially for thoracic tumors affected by respiratory motion.
Purpose of the Study:
- To develop and evaluate an innovative Mask-based Swin Transformer network (MBST) for enhancing 4D-CBCT reconstruction quality.
- To assess the MBST's performance across a wide range of scanning conditions, including limited projection data and high scanning speeds.
Main Methods:
- A deep learning model, MBST, was trained using 4D-CBCT data from 20 thoracic tumor patients, with simulated 4D-CBCT generated from 4D-CT data.
- The network was trained on data reconstructed under limited scanning conditions and tested on various projection intervals (1° to 24°) and 1/3 full angles.
- Quantitative evaluation used SSIM, PSNR, ME, and MAE; qualitative assessment and comparisons with other deep learning methods were performed on both simulated and real clinical data.
Main Results:
- MBST significantly improved 4D-CBCT image quality, showing a 42.3% increase in SSIM and 10.8 dB in PSNR at a 4° interval, with ME and MAE approaching zero.
- The network demonstrated superior performance compared to other deep learning methods, achieving higher SSIM and PSNR, and lower MAE across different projection intervals.
- MBST successfully reconstructed bone tissue, lung structures, and tracheal walls, and significantly improved clinical 4D-CBCT quality metrics (SSIM, PSNR, ME, MAE) compared to 4D-CT registration.
Conclusions:
- The Mask-based Swin Transformer network (MBST) effectively reconstructs high-quality 4D-CBCT images under diverse scanning conditions, including limited data acquisition.
- MBST demonstrates excellent generalization capabilities and achieves satisfactory CT values and image quality on clinical patient datasets.
- MBST serves as a highly generalized reconstruction network, offering a significant advancement for improving 4D-CBCT image quality in clinical practice.
