Author: Hideaki Hirashima, Haruo Inokuchi, Nobutaka Mukumoto, Naruki Murahashi, Mitsuhiro Nakamura, Megumi Nakao, Keiko Shibuya, Linna Zhang 👨🔬
Affiliation: Kyoto University, Osaka Metropolitan University 🌍
Purpose:
To transform the quality of 2D cine MR images acquired during online adaptive MR-guided radiotherapy (OA-MRgRT) by utilizing a conditional diffusion model to achieve image quality comparable to that of T2-weighted 3D MR images.
Methods:
MR images were collected from 38 prostate cancer patients with a peri-SpaceOAR abscess who underwent OA-MRgRT with the Elekta Unity MR-linac, comprising a total of 398 treatment fractions. Each 2D axial cine image acquired during beam delivery (2Dori) was paired and aligned with a corresponding 2D axial slice extracted from a pre-treatment T2-weighted 3D MR image (2DGT) through rigid registration. Each axial image matrix was standardized to a size of 256×256 with an isotropic resolution of 0.83 mm. The 99th percentile value (P99) of pixel intensity in each slice was set as the threshold, after which the pixel intensity values were normalized to the range of [-1, 1] from [0, P99]. A total of 3,020 paired 2D images were used to train the conditional diffusion model. Finally, the model was tested on an additional 553 2D axial slices, and the differences in similarity between the transformed 2D images (2Dtrans) and 2DGT, and between 2Dori and 2DGT, were assessed.
Results:
The testing produced 2Dtrans with the median normalized cross-correlation (NCC) of 0.91 (IQR: 0.88–0.92) and the structural similarity index measure (SSIM) of 0.83 (IQR: 0.73–0.87) relative to the corresponding 2DGT. This represents an improvement from the median NCC value of 0.63 (IQR: 0.59–0.69) and the SSIM value of 0.36 (IQR: 0.27–0.45) calculated from 2Dori and the corresponding 2DGT. The conditional diffusion model demonstrated significant improvements in NCC and SSIM (p < 0.05).
Conclusion:
Utilizing the conditional diffusion model, 2D MR images acquired during OA-MRgRT can be efficiently transformed into high-quality images comparable to those of T2-weighted MR images without artifacts.