dtiRIM: A generalisable deep learning method for diffusion tensor imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Diffusion weighted MRI is crucial for diagnosing pathologies.
• Current deep learning methods for diffusion parameter quantification lack generalizability across different scan protocols.
• This necessitates retraining neural networks for each new dataset, limiting clinical adoption.

Purpose of the Study:

• To introduce dtiRIM, a novel deep learning method for Diffusion Tensor Imaging (DTI).
• To develop a generalizable DTI analysis tool that overcomes the limitations of existing deep learning approaches.
• To enable high-quality tensor estimation across diverse acquisition settings with a single trained model.

Main Methods:

• dtiRIM utilizes Recurrent Inference Machines to solve inverse problems in DTI.
• The method incorporates the diffusion tensor model to ensure data consistency.
• Validation involved simulations and in vivo data, comparing dtiRIM against Iterated Weighted Linear Least Squares (IWLLS) and Maximum Likelihood Estimator (MLE).

Main Results:

• dtiRIM demonstrated low dependency on tissue properties, anatomy, and scanning parameters.
• Performance was comparable to or better than established methods like IWLLS and MLE.
• A single dtiRIM model successfully processed diverse datasets without significant quality degradation.

Conclusions:

• dtiRIM represents a significant advancement in DTI analysis, offering unprecedented generalizability.
• The method provides a robust and adaptable solution for quantifying diffusion parameters from MRI data.
• This work presents the first generalizable deep learning-based solver for DTI, paving the way for broader clinical application.