Report on the quantitative intra-voxel incoherent motion diffusion MRI reconstruction grand challenge

Area of Science:

• Medical Imaging
• Biophysics
• Computational Biology

Background:

• Diffusion MRI (dMRI) is crucial for non-invasive tissue characterization.
• Quantitative parameter estimation in dMRI, particularly using the IVIM model, faces challenges in accuracy and robustness.
• Clinical adoption of dMRI is hindered by limitations in current reconstruction techniques.

Purpose of the Study:

• To benchmark and advance reconstruction algorithms for quantitative IVIM-dMRI.
• To improve the accuracy and robustness of IVIM parameter estimation for clinical applications.
• To address barriers hindering the broader clinical adoption of IVIM-dMRI.

Main Methods:

• Participants reconstructed IVIM parameters (fractional perfusion, pseudo-diffusion, true diffusion coefficients) from simulated k-space data.
• The challenge utilized realistic digital VICTRE phantoms across training, validation, and testing phases.
• Performance was evaluated using relative root mean square error (rRMSE); both traditional and deep learning (DL) methods were permitted.

Main Results:

• 42 teams participated, with 7 advancing to the final phase, demonstrating broad engagement.
• The relative root mean square error (rRMSE) varied between 0.0345 and 1.24.
• A cascaded U-Net architecture achieved top performance, highlighting the potential of DL in complex medical imaging inverse problems.

Conclusions:

• The IVIM-dMRI challenge successfully advanced reconstruction accuracy and robustness.
• Deep learning approaches show significant potential for improving IVIM parameter estimation.
• Future work should focus on real-world data complexities to enhance clinical applicability.