Quiz: Physics & Engineering: Insights into Learning-Based MRI Reconstruction

Which recent changes sparked the recent rise of machine learning applications?
- Big data, increased computational power, and algorithmic development
- Unhappiness with compressed sensing reconstruction quality and computational burden
- Increased access to training data to due to relaxed data privacy regulations (e.g., HIPAA)
Jump to the relevant part of the video for the above question:
What is a key difference between compressed sensing and machine learning reconstructions?
- Compressed sensing relies on nonlinear optimization methods whereas learning requires no iterative optimization at any stage.
- Compressed sensing reconstruction can produce images with residual artifacts or quality loss, while machine learning will always generate diagnostic quality images.
- In compressed sensing sparsity transforms, penalties, and parameters are hand-selected whereas in machine learning they are automatically determined from training data
Jump to the relevant part of the video for the above question:
Within which stage of machine learning reconstruction does most of the computational burden lie?
- Training the network from reference data
- Generating a image from new data using pre-trained network
- Both 1 and 2 take require approximately the same amount of computation
Jump to the relevant part of the video for the above question:
Which of the following statements about the training data used in machine learning reconstruction is false:
- A sufficiently large number of training examples is needed to avoid network overfitting
- Since machine learning MRI reconstruction operates on k-space data, artifacts present in the training images will not be learned
- Different types of MR images (e.g., single vs. multiple coil) can potentially be used for training
Jump to the relevant part of the video for the above question:
Once a machine learning reconstruction network is trained, it will typically generate high-quality images for which type of new data:
- Any MRI data which is input to the network
- Any MRI data acquired using a similar protocol and which is anatomically/pathologically similar to the training data.
- Any MRI data of anatomy which was included in the training set, even if pathologically different
Jump to the relevant part of the video for the above question: