L Kong¹*, Z Li², Q Zhou¹, D Huang³, Y Yin², (1) Manteia Medical Technologies, Milwaukee (2) Department of Radiation Oncology Physics and Technology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Ji'nan, Shandong province, CN, (3) College Of Engineering, Huaqiao University, Quanzhou

• 
  L Kong

PO-GePV-E-21 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Multi-modality image registration plays an important role in disease diagnosis and staging, tumor delineation, and prognostic analysis. However, it is often challenging due to the distinct image variations among different modalities. In this study, we proposed a synthesis-based unsupervised multi-modality registration network--Fork-Net.

Methods: In Fork-Net, the generator, discriminator and registration network were integrated into a single model with a reused encoder. The moving image was first synthesized into the same mode as the fixed image, and then registered through registration network, turning the multi-modality registration task into a mono-modality registration problem. A new Synthesis-Registration-Consistency loss was thus developed to strengthen the consistency constraint between generator and registration network. The method was demonstrated using two brain datasets: one open access (BraTS, 2018) for T1-T2 MRI registration and one in-house dataset for MR-CT registration. The model performance was compared with the state-of-the-art image synthesis and registration methods. The quality of the image synthesis was evaluated using mean absolute error (MAE) and larger peak signal-to-noise ratio (PSNR). The registration accuracy was evaluated through the Dice of the target contour before and after registration, and the average gradient of deformation field (aGDF).

Results: Regarding synthetic image quality, Fork-net has smaller MAE and larger PSNR. The target contours using registered images through ForkNet has the highest Dice, i.e., 0.74 ±0.12 and 0.68±0.14 for dataset 1 and 2, respectively. ForkNet has the smallest aGDF, indicating smoothest deformation. In addition, the architecture of Fork-Net is superior to other models with less adjustable parameters and less memory requirement.

Conclusion: The proposed synthesis-based unsupervised multi-modality registration network is superior in terms of synthesis quality, registration accuracy, and smoothness of the deformation field as compared with current state-of-the-art methods. It may be implemented into clinical workflow, facilitating routine practice of multi-modality image registration.

ePosters

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email