MoCo-SToRM
In this blog, I would like to show some results, especially some exciting movies that I am not showing in the paper: Dynamic imaging using Motion-Compensated SmooThness Regularization on Manifolds (MoCo-SToRM).
MoCo-SToRM is an unsupervised motion-compensated reconstruction scheme using smoothness regularization on manifolds for high-resolution free-breathing pulmonary MRI. Basically, we model the image frames in the time series as the deformed version of the 3D template image volume. We assume that the deformation maps to be points on a smooth manifold in high-dimensional space. Specifically, we model the deformation map at each time instant as the output of a CNN based generator that is has the same weights for all time-frames, driven by a low-dimensional latent vector. The time series of latent vectors account for the dynamics in the dataset including respiratory motion and bulk motion. The template image volume, the parameters of the generator, and the latent vectors are learned directly from the k-t space data in an unsupervised fashion.
This MoCo-SToRM scheme is pretty robust to the bulk motion in the dataset because we use the learnable latent vectors to capture the dynamics in the images. This offer the latent vectors the ability to detect both usual motion and bulk motion. In this post, we will show some movies about the two datasets, which are corrupted by bulk motions. One dataset (post-contrast) was acquired on a adult subject who was diagnosed as pulmonary fibrosis. Another dataset is acquired from a neonatal subject who was admitted into the neonatal intensive care unit (NICU) at Cincinnati children’s hospital because of severe bronchopulmonary dysplasia. More information about the datasets can be found in the paper.
Both these two datasets are associated with bulk motions, especially the NICU dataset, in which extensive bulk motions can be detected.
We first show some results from the dataset acquired from the adult subject.
Movie 1. This movie show 100 frames from the coronal view. During this period, no bulk motion happens and the subject breathes normally. We also show the corresponding time profiles and the latent vectors. From which we can see that when the patient breathes normally, the time profiles and the latent vectors are pretty normal.
Movie 2. This movie show 100 frames from the coronal view. During this period, we can detect one bulk motion except for the usual respiratory motion. We show the corresponding time profiles and the latent vectors. From which we can see that when the bulk motion happens, there will be sudden jumps in both the time profiles and the latent vectors. If we look at the movie, we can clearly see that when the sudden jump happens, the shoulder on the left side moves.
Next, we show some results from the dataset acquired from the neonatal subject. Unlike the dataset acquired from the adult subject, which only has one bulk motion. This dataset contains lots of bulk motion as this dataset is acquired from a neonatal subject and it is hard to let them stay still in the MR scanner.
Movie 3. This movie show one cut of the whole scan using the axial view. From which we can see that the baby moves a lot during the scan.
Movie 4. This movie show one cut (same as the cut used in Movie 3) of the whole scan using the coronal view. From which we can see that the baby moves a lot during the scan.
