Self-supervised IVIM DWI parameter estimation with a physics based forward model.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Diffusion-weighted MRI (DW-MRI) is crucial for abdominal organ assessment.
• Intravoxel Incoherent Motion (IVIM) modeling provides insights into tissue microdynamics.
• Noisy signals in DW-MRI can compromise IVIM parameter estimation accuracy.

Purpose of the Study:

• To evaluate a novel neural network method for robust IVIM parameter estimation in abdominal organs.
• To assess the repeatability of the proposed method under noisy diffusion signal conditions.
• To compare the novel method against conventional non-linear least squares (NNLS) and artificial neural network (ANN) approaches.

Main Methods:

• Retrospective analysis of clinically acquired abdominal DW-MRI scans from Crohn's disease patients.
• Segmentation of regions in kidney cortex, spleen, liver, and bowel.
• Application of a physics-guided self-supervised convolutional neural network for IVIM fitting.
• Comparison with NNLS and voxelwise trained ANN methods.

Main Results:

• The novel neural network method demonstrated increased robustness to noise.
• Improved intra-session repeatability with reduced coefficient of variation (CoV) compared to NNLS.
• Comparable repeatability to ANN; superior performance in classifying normal vs. abnormal Crohn's disease bowel tissue.
• Challenges remain in fitting the $$ parameter.

Conclusions:

• The proposed method provides robust D and f IVIM parameter estimates from noisy diffusion signals.
• This technique holds promise for accelerated DW-MRI acquisition strategies.
• Potential for improved diagnostic accuracy in conditions like Crohn's disease.