Cross-modality brain tumor MRI image synthesis method based on attention enhancement and edge awareness(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2026年第1期
- Page:
- 65-75
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Cross-modality brain tumor MRI image synthesis method based on attention enhancement and edge awareness
- Author(s):
- LI Hao1; 2; YANG Zhihui1; 2; LI Fengsen1; 2
- 1. School of Medical Information Science and Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China 2. Gansu Provincial Hospital/First School of Clinical Medicine, Gansu University of Chinese Medicine, Lanzhou 730000, China
- Keywords:
- generative adversarial network brain tumor MRI image synthesis attention mechanism edge awareness gradient regularization
- PACS:
- R318
- DOI:
- DOI:10.3969/j.issn.1005-202X.2026.01.009
- Abstract:
- Objective To develop a high-quality cross-modality synthesis method for magnetic resonance imaging (MRI) of brain tumor, tackling the challenges of high acquisition time cost, multiple artifacts, and incomplete modality availability in MRI. Methods A registration-enhanced generative adversarial network named AE?egGAN integrating attention mechanism and edge awareness was developed for cross-modality synthesis from T1-weighted to T2-weighted MRI images. In the generator, a CoordAttention module was embedded to enhance the perception of key regions, while a Sobel-based edge detection mechanism was incorporated to strengthen the representation of tumor boundaries. In the discriminator, gradient penalty regularization was adopted to improve training stability and mitigate mode collapse. Results The model was trained on 5 760 brain tumor MRI samples and evaluated on 768 test cases. AE?egGAN outperformed the original RegGAN, exhibiting an improvement of 0.51 dB in peak signal-to-noise ratio (PSNR) and 0.029 in structural similarity index measure (SSIM) in local tumor regions. For global image, the PSNR and SSIM were enhanced by 0.900 dB and 0.032, respectively. Paired t?est results for global image demonstrated statistically significant improvements in mean absolute error (P=0.026 4), PSNR (P<0.000 1), and SSIM (P<0.000 1). Ablation experiments further validated the effectiveness of the attention and edge awareness modules. Conclusion AE?egGAN exhibits superior structural preservation capability and lesion sensitivity in multimodal brain MRI synthesis, offering a stable and reliable image completion solution to support clinical diagnosis.
Last Update: 2026-01-27