L Kong¹*, W Jiang², D Huang¹, Q Zhou³, Z Chen³, W Zhang³, H Sheng³, C Yang³, B Qu⁴, (1) College Of Engineering, Huaqiao University, Quanzhou (2)Department of Radiotherapy, Yantai Yuhuangding Hospital, Qingdao University School of Medicine, Yantai, Shandong, 264000,China. (3) Manteia Medical Technologies (4) PLA General Hospital, Beijing, ,CN

• 
  L Kong

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-E-20 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Existing work on unpaired MR-to-CT image synthesis performs well at the 2D level. However, there is still some limitations regarding the synthetic image quality in the three dimensions. For instance, there is a lack of continuity between the previous and subsequent layers of the 3D sCT image, which is manifested in uneven levels in tissues and bones. In addition, the HU value of the 3D synthetic CT image has minor variations, indicating losing structural details in coronal and sagittal planes. Moreover, there is lack of shape consistency with the original MR. In order to solve the above problems, an effective spatial sequence generative adverse network (ssgan) was proposed to achieve unpaired MR to CT image synthesis.

Methods: The convolutional long short-term memory (Conv-LSTM) model was embedded in the generator to guarantee the 3D spatial continuity of synthesis images by using a mixture of the input 2D images in different spatial orders. Secondly, three projection discriminators were introduced to discriminate the authenticity of the three projected images along axial, coronal and sagittal planes. Moreover, a non-local module was incorporated into the discriminator to further enhance the strength of distinguishing the voxel intensity relationship in the adjacent layers.

Results: The feasibility and effectiveness of our SSGAN was demonstrated using unpaired head and neck MR and CT images. The synthetic CT (sCT) generated using our SSGAN shows superior similarity to the real CT in both 3D continuity and intensity authenticity. In addition, the cross-modality registration experiment illustrates that our SSGAN generates better sCT with highest mutual information

Conclusion: the novel SSGAN method can provide better quality 3D sCT images from unpaired MR data compared with others methods.

ePosters

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email