基于改进U-Net模型的脑部MRI生成合成CT研究

Improved U-Net model for generating synthetic CT from brain MRI

摘要

目的：提出一种基于改进U-Net的脑部MRI生成合成CT（sCT）的方法，实现MRI到CT图像的转换，并比较传统U-Net模型及改进模型在生成脑部sCT方面的性能差异。方法：基于用于放疗计划的CT图像生成2023挑战赛（SynthRAD2023）数据集，选取120例患者的脑部MRI、CT图像作为训练集，39例作为测试集。分别构建传统U-Net模型以及基于U-Net改进的模型。对于二者生成的sCT，分别从图像质量、强度值等角度进行评估。结果：对于测试集中的病例，使用传统U-Net模型和改进模型生成的sCT图像与真实CT间的平均绝对误差分别为（100.28±49.67）HU和（94.87±42.50）HU，峰值信噪比分别为（26.43±2.70）dB和（26.85±2.73）dB，结构相似性分别为0.808±0.092和0.820±0.085。结论：本研究基于U-Net改进的模型在脑部MRI生成sCT精度方面优于传统的U-Net模型。

Abstract

Objective To propose an improved U-Net approach for generating synthetic computed tomography (sCT) from brain magnetic resonance imaging (MRI), achieving MRI-to-sCT translation, and further compare the brain sCT generation performance of the improved model with that of the traditional U-Net model. Methods The SynthRAD2023 dataset for CT image generation in radiotherapy planning was adopted in the study. The brain MRI and CT images of 120 patients were selected as the training set, while those from 39 patients were used as the test set. A traditional U-Net model and an improved U-Net model were constructed, and their generated sCT were evaluated regarding image quality and intensity values. Results For the cases in the test set, the average absolute error relative to ground truth CT reached (100.28±49.67) HU for the traditional U-Net model and (94.87±42.50) HU for the improved model. The corresponding peak signal-to-noise ratios were (26.43±2.70) dB and (26.85±2.73) dB, and the structural similarities were 0.808±0.092 and 0.820±0.085, respectively. Conclusion The proposed improved U-Net model outperforms the traditional U-Net model in terms of accuracy for brain MRI-to-sCT translation.

关键词

磁共振成像 / 脑 / 合成CT / U-Net

Key words

magnetic resonance imaging / brain / synthetic computed tomography / U-Net

参考文献

[1] Simkó A, Bylund M, Jönsson G, et al. Towards MR contrast independent synthetic CT generation[J]. Z Med Phys, 2024, 34(2): 270-277.

[2] Chourak H, Barateau A, Tahri S, et al. Quality assurance for MRI—only radiation therapy: a voxel—wise population—based methodology for image and dose assessment of synthetic CT generation methods[J]. Front Oncol, 2022, 12: 968689.

[3] Lerner M, Medin J, Jamtheim Gustafsson C, et al. Prospective clinical feasibility study for MRI—only brain radiotherapy[J]. Front Oncol, 2022, 11: 812643.

[4] Yang X, Feng B, Yang H, et al. CNN—based multi—modal radiomics analysis of pseudo—CT utilization in MRI—only brain stereotactic radiotherapy: a feasibility study[J]. BMC Cancer, 2024, 24(1): 59.

[5] Young T, Dowling J, Rai R, et al. Clinical validation of MR imaging time reduction for substitute/synthetic CT generation for prostate MRI—only treatment planning[J]. Phys Eng Sci Med, 2023, 46(3): 1015-1021.

[6] 李芳华, 丁寿亮, 李永宝, 等 . 基于Transformer的生成对抗网络用于鼻咽癌MRI生成伪CT[J]. 中国医学物理学杂志, 2025, 42(6): 701-707.

[7] Li FH, Ding SL, Li YB, et al. Synthetic CT generation from NPC MRI using Transformer—based generative adversarial network[J]. Chinese Journal of Medical Physics, 2025, 42(6): 701-707.

[8] Hoffmann A, Oborn B, Moteabbed M, et al. MR—guided proton therapy: a review and a preview[J]. Radiat Oncol, 2020, 15(1): 129.

[9] Albertini F, Matter M, Nenoff L, et al. Online daily adaptive proton therapy[J]. Br J Radiol, 2020, 93(1107): 20190594.

[10] Li X, Bellotti R, Meier G, et al. Uncertainty—aware MR—based CT synthesis for robust proton therapy planning of brain tumour[J]. Radiother Oncol, 2024, 191: 110056.

[11] 邵虹, 荆一烜, 崔文成 . 基于扩散生成对抗网络的核磁共振图像与计算机断层扫描图像跨模态转换[J]. 生物医学工程学杂志, 2025, 42(3): 575-584.

[12] Shao H, Jing YX, Cui WC . Cross modal translation of magnetic resonance imaging and computed tomography images based on diffusion generative adversarial networks[J]. Journal of Biomedical Engineering, 2025, 42(3): 575-584.

[13] Thummerer A, van der Bijl E, Galapon A Jr, et al. SynthRAD2023 grand challenge dataset: generating synthetic CT for radiotherapy[J]. Med Phys, 2023, 50(7): 4664-4674.

[14] El Jurdi R, Petitjean C, Honeine P, et al. CoordConv—Unet: investigating CoordConv for organ segmentation[J]. IRBM, 2021, 42(6): 415-423.

[15] Liu MK, Zhuo YT, Liu J, et al. Real—time estimation of lung deformation from body surface using a general CoordConv CNN[J]. Comput Methods Programs Biomed, 2024, 244: 107998.

[16] Yin AC, Ren C, Yue WT, et al. CDAU—Net: a novel CoordConv—integrated deep dual cross attention mechanism for enhanced road extraction in remote sensing imagery[J]. Remote Sens (Basel), 2023, 15(20): 4914.

[17] Fan XS, Ding WT, Qin WL, et al. Fusing self—attention and CoordConv to improve the YOLOv5s algorithm for infrared weak target detection[J]. Sensors (Basel), 2023, 23(15): 6755.

[18] Liu Z, Mao HZ, Wu CY, et al. A ConvNet for the 2020s[C]// 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Piscataway, NJ, USA: IEEE, 2022: 11966-11976.

[19] Song JH, Nie XB, Wu C, et al. A novel intelligent fault diagnosis method of rolling bearings based on the ConvNeXt network with improved DenseBlock[J]. Sensors (Basel), 2024, 24(24): 7909.

[20] Liu SK, Cao S, Lu X, et al. Lightweight deep learning model, ConvNeXt—U: an improved U—Net network for extracting cropland in complex landscapes from Gaofen—2 images[J]. Sensors (Basel), 2025, 25(1): 261.

[21] Wei FM, Wang X . Sar ship detection based on convnext with multi—pooling channel attention and feature intensification pyramid network[J]. Sensors (Basel), 2023, 23(17): 7641.

[22] Dai JF, Qi HZ, Xiong YW, et al. Deformable convolutional networks[C]// 2017 IEEE International Conference on Computer Vision (ICCV). Piscataway, NJ, USA: IEEE, 2017: 764-773.

[23] Misra D, Nalamada T, Arasanipalai AU, et al. Rotate to attend: convolutional triplet attention module[C]// 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). Piscataway, NJ, USA: IEEE, 2021: 3138-3147.

[24] Tang L, Yi JZ, Li XY . Improved multi—scale inverse bottleneck residual network based on triplet parallel attention for apple leaf disease identification[J]. J Integr Agric, 2024, 23(3): 901-922.

[25] Li Y, Yan BB, Hou JX, et al. UNet based on dynamic convolution decomposition and triplet attention[J]. Sci Rep, 2024, 14(1): 271.

[26] Li XX, Xue SY, Xie JY, et al. Interactive triplet attention for few—shot fine—grained image classification[J]. Neurocomputing, 2025, 655: 131377.

[27] Meng WJ, Shan LL, Ma SG, et al. DLNet: a dual—level network with self—and cross—attention for high—resolution remote sensing segmentation[J]. Remote Sens (Basel), 2025, 17(7): 1119.

[28] Li FD, Lu XY, Yuan JJ . MHA—CoroCapsule: multi—head attention routing—based capsule network for COVID—19 chest X—ray image classification[J]. IEEE Trans Med Imaging, 2022, 41(5): 1208-1218.

[29] Wang Z, Bovik AC, Sheikh HR, et al. Image quality assessment: from error visibility to structural similarity[J]. IEEE Trans Image Process, 2004, 13(4): 600-612.

基金资助

甘肃省科技计划项目-软科学专项(25JRZA204)