Focus Session

SCMR 22nd Annual Scientific Sessions

6 - DeformationNet: unsupervised deep learning co-registration for ventricular segmentation of short axis perfusion MR images

Presenter(s)

• BN

  Berk Norman

  Machine Learning Scientist
  

Author Only(s)

• MS

  Martin Simonovsky

  Machine Learning Intern
  Arterys, Inc.
• FL

  Felix Lau, BEng

  Senior Machine Learning Scientist
  Arterys, Inc.
• SS

  Sean Sall, BA

  Senior ML Scientist
  Arterys
• DG

  Daniel Golden, PhD

  Director of Machine Learning
  Arterys, Inc
• AH

  Albert Hsiao, MD, PhD

  Assistant Professor
  University of California, San Diego

Background: Cardiac perfusion MR imaging is frequently performed to identify perfusion defects that correlate with any coronary vascular distributions. Semi-quantitative evaluation of these perfusion defects requires segmentation of the myocardium over multiple image frames to compute perfusion characteristics on a per-pixel or per-myocardial segment basis. As such, automatic segmentations of the left ventricle endocardium (LV endo) and epicardium (LV epi) may be essential tools to improve reproducibility and efficiency of interpretation.  As such, we propose a deep learning framework, DeformationNet, which allows for unsupervised image co-registration of perfusion MR images and semi-supervised ventricular segmentation mask capable of transferring across multiple time points.

Methods: Using a previously trained ventricular segmentation network that was trained on cine SSFP images, inference is  first run on all time points in a perfusion scan. A heuristic (depicted in Figure 1), based on the inferred segmentation probability maps, is then used to select the perfusion image with the “best” target segmentation mask for the given slice.

DeformationNet (depicted in Figure 2) is trained to predict dense deformation fields (DDF) between two perfusion MR images in order to register the moving image onto a fixed image. The DDF is optimized by minimizing the mutual information difference between the fixed image and the DDF applied to the moving image.

To use the trained DeformationNet to transfer ventricular segmentation masks between time points, the best target perfusion slice, selected by the heuristic, serves as the moving image for DeformationNet, and all other slices are fed in (one at a time) as the fixed image to generate a DDF. The respective DDFs are then applied to the target segmentation mask to generate a warped segmentation mask for each fixed image.

To evaluate the performance of this mask transferring, manual ventricular segmentations were made by expert annotators for a 42 scan hold-out test set and dice scores were computed between these annotations and the transferred mask segmentations. The demographic distribution of these sets can be viewed in Table 1.

Results: Median Dice values on the test set were 0.855 (IQR values: 0.698-0.899) for LV endo and 0.879 (IQR values: 0.715-0.925) for LV epi. Average computation for calculating the DDFs for an entire perfusion volume was approximately 7 seconds on a Nvidia P100 GPU (compared to traditional iterative co-registration methods which can take in excess of one hour).

Conclusion: We propose DeformationNet, an approach using unsupervised neural networks that can quickly produce deformation fields between new sets of images. These can be leveraged to perform image co-registration and transferring of segmentation masks or anatomical landmarks across multiple image frames. This approach may be faster and more accurate than traditional iterative registration methods and can greatly improve the workflow of perfusion analysis.