303.1 Semester Project Presentation

What Is DWI ?

MRI technique that measures the diffusion of water molecules in biological tissues.

• Principle: Uses magnetic field gradients to sensitize the MRI signal to water diffusion
• Key parameters:
  • b-values
  • b-vectors

Figure 1: Visual example of DWI, sourced from Dr. A. Rokem, DIPY workshop 2021

Why DWI ?

Figure 2: NDMG detailed pipeline

Project Goals

• Contribute to DIPY
• Add prediction capabilities to
  the existing GeneralizedQSamplingModel class

What Has Been Done

• Enabled prediction for the GeneralizedQSamplingModel class
• Rigorous Testing
  • Unit tests (100% coverage)
  • Integration tests via different Marimo notebooks
• Automated code documentation via Sphinx
• Created a DIPY tutorial document on the prediction pipeline
• Project documentation (see homepage)
• Ongoing pull request on Prof.Estban DIPY fork
  • Currently being reviewed by the DIPY’s team

DIPY GQI Model Pipeline

Figure 3: DIPY GQI model simplified pipeline, made with GraphViz

Figure 4: GQI model input image example,
from the stanford_hardi dataset¹

→

Figure 5: GQI model output example (ODF),
from the stanford_hardi dataset¹

1. The stanford_hardi dataset is a high angular resolution diffusion imaging (HARDI) dataset from Stanford University, consisting of 150 diffusion-weighted volumes.

How Is The Signal Predicted ?

\[ \begin{alignat*}{2} &ODF &&= S \cdot K \\ \iff\quad & ODF \cdot K^{+} &&\approx S \cdot K \cdot K^{+} \\ \iff\quad & ODF \cdot K^{+} &&\approx S \\ \iff\quad & S &&\approx ODF \cdot K^{+} \end{alignat*} \]

Where:

• \(S\) is the GQI model fitted signal (or data)
• \(K^{+}\) is a regularized pseudo-inverse of the GQI kernel \(K\).

Testing Methodology

• Signal reconstruction based on seen data
  • To make sure that the predict function implementation works properly
• Signal prediction based on unseen data
  • To test the generalization of the predict function

Testing Methodology: Reconstruction Tests

Figure 6: Reconstruction pipeline schema, made with excalidraw

Single-Voxel Reconstruction Results

Figure 7: Voxel (0, 0, 0) Signal Reconstruction: Original vs Predicted based on the small_101D dataset¹

1. The small_101D dataset is a small subset of diffusion spectrum imaging (DSI) data from DIPY, containing 101 diffusion-weighted volumes.

Multi-Voxel Reconstruction Results

Figure 8: Voxel-wise Reconstruction Accuracy based on the small101D dataset

Testing Methodology: Unseen Prediction Test

Figure 9: Unseen data prediction pipeline, made with excalidraw

Unseen Prediction Results

Figure 10: Unseen data prediction results based on the taiwan_ntu_dsi dataset¹

1. The taiwan_ntu_dsi dataset is a full DSI acquisition from the National Taiwan University, consisting of 515 diffusion-weighted volumes.

Comparison Of
Reconstruction Vs Unseen Predictions

Figure 11: Comparison of reconstruction vs unseen predictions based on the taiwan_ntu_dsi dataset

Impact

• Head motion noise reduction (see Ben-Amitay, Jones, and Assaf (2012))
• Meaningful data augmentation
• Model cross validation

Conclusion

• Great introduction to medical data science
• Interesting to work on an established open source project
• Useful bachelor thesis warm up

Additional Resources

If you want more information about the project please visit this project website
https://olivier.amacker.dev/303.1-semester-project/

Figure 12: Project website QR code

References

Ben-Amitay, Shani, Derek K. Jones, and Yaniv Assaf. 2012. “Motion Correction and Registration of High b-Value Diffusion Weighted Images.” Magnetic Resonance in Medicine 67 (6): 1694–1702. https://doi.org/10.1002/mrm.23186.