Author: Nan Li, Shouping Xu, Gaolong Zhang, Xuerong Zhang 👨🔬
Affiliation: Department of Radiation Oncology, HeBei YiZhou proton center, School of Physics, Beihang University, National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College 🌍
Purpose:
The 3D BRAVO sequence is an advanced magnetic resonance (MR) technique that allows for image reconstruction at any angle. It offers 1 mm gapless scanning and has a high signal-to-noise ratio, making it particularly beneficial for detecting small tumors. In radiation therapy (RT), 3D BRAVO images assist doctors in delineating target volume more accurately than CT. However, 3D BRAVO scanning is costly, not widely available in hospitals, and unsuitable for patients with claustrophobia or metal implants. Using deep learning to convert CT images into 3D BRAVO images can reduce time and costs, aid doctors in delineating RT target areas, and provide an alternative for patients unable to undergo scanning.
Methods:
We propose an innovative network architecture, AGC-GAN. In the outermost layer of the generator, we introduce a spatial attention block, while the middle layer of the discriminator employ a channel attention block. Furthermore, we incorporate a grayscale histogram loss function during training to ensure that the image enhancement process considers global features and local details. Our experimental dataset consists of CT images from 173 patients, with 130 cases used for model training and the remaining 43 cases for model testing.
Results:
The model-generated 3D BRAVO MR images are highly similar to the real 3D BRAVO MR scans. Specifically, the peak signal-to-noise ratio (PSNR) has an average value of 32.21 dB±3.51; the structural similarity index (SSIM) averages 0.97±0.02; and the normalized mean squared error (NMSE) has an average of 0.04±0.02.
Conclusion:
Experimental results demonstrate that the proposed AGC-GAN can efficiently and accurately generate pseudo 3D BRAVO MR images from CT. The generated images exhibit high consistency with the real scan, enabling their use in image fusion to assist radiation oncologists in delineating target regions, thereby comprehensively optimizing the radiotherapy workflow.