Synopsis

Keywords: Image acquisition: Reconstruction, Transferable skills: Software engineering

This presentation will give an overview on the practical aspects of implementing forward and backward operators for MR image reconstruction. We will cover the high-level details of implementing non-Cartesian compressed sensing reconstruction in PyTorch. Topics covered will include 1) how to convert mathematical operators to code, 2) code execution on the CPU vs. GPU, 3) building linear operations and verifying their correctness, and 4) a final demonstration of reconstruction. Although the implementation focuses on PyTorch, high-level concepts will be extensible to other languages such as MATLAB or Julia, and references will be made to these alternative frameworks where possible.

Summary

The objective of this presentation is to give an overview on the practical aspects of implementing forward (and backward) operators for MR image reconstruction. Although most methods are first developed with general mathematical frameworks and notation, at some point they must be implemented on a computer. We will cover the high-level details of implementing a compressed sensing reconstruction in PyTorch. Although we focus on PyTorch, most modern languages/frameworks used for computational imaging (such as MATLAB, Julia or Numba) will include the features used in the implementation.

In particular we will cover how to structure an operator in practical code. This will include first understanding the target operation’s mathematical details (in our case a non-Cartesian SENSE forward operator), linking those details to the appropriate software library, and then implementing them in a forward operation. After implementing the forward operation, it is also necessary to implement the backward operation and verify that the backward operation satisfies the adjoint properties. We will then apply the same procedure to any regularization operations. After wrapping up by setting up the pre-processing code, we will demonstrate non-Cartesian, compressed-sensing regularized reconstruction on the CPU.

After initial implementation, we will focus on more advanced topics to increase speed. This will include GPU-based reconstruction, coil/slice broadcasting. The presentation will also discuss some of the internals of the library operations in use to demonstrate what aspects are likely to increase their computation speed, and the presentation will conclude with a reconstruction demonstration with improved speed.

Acknowledgements

No acknowledgement found.