R Wei*, B Liang, K Men, J Dai, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, BeijingCN,

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-I-28 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: The 3D-MRI reconstruction from single cine-MRI has a great value for estimating the tumor motion in 3D space for MR-gRT. However, the current methods utilized only personal or population data as prior knowledge for 3D-MRI reconstruction, which might compromise the accuracy and robustness. Thus, we proposed a transfer-learning-based 3D-MRI reconstruction method from single cine-MRI, which combined the personal and population prior knowledge.

Methods: In this method, a deep learning network that had a two-branch structure for input was trained with 4D-MRI from 10 patients. The network took 2D MRI slice and a reference 3D-MRI as the input, and directly reconstructed 3D-MRI corresponded to the 2D MRI slice. Then, for the unseen patients, 4D-MRI was acquired before the treatment. The network was fine-tuned with 2D MRI slices and reference 3D MRI that were generated from untreated patients’ 4D-MRI. Finally, the network could be utilized for 3D-MRI reconstruction by using cine-MRI and reference 3D-MRI as the input for the unseen patients.

Results: We utilized 4D-MRI from 5 patients to test the proposed method. For each test patient, 3D-MRI of phase 0%-30% were utilized for network fine-tuning and 3D-MRI of phase 40%-90% were used for evaluation. The tumor localization error in 3D space was utilized as evaluation metrics. Moreover, we compared the proposed method with the personal prior knowledge based method and population prior knowledge based method. Due to the irregular breathing, the tumor localization errors for personal knowledge based method were over 7.9 mm±3.3mm. Meanwhile, the tumor localization errors of population knowledge based method were below 3.5mm±1.6mm. As for the proposed method, it had the best performance, whose tumor localization errors were under 2.0 mm±1.1 mm.

Conclusion: We proposed a method for 3D MRI reconstruction from a single cine-MRI image using transfer learning, which can achieve robust and accurate 3D MRI reconstruction

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email