Accurate estimation of intravoxel incoherent motion parameters based on implicit neural representation

Area of Science:

• Medical Imaging
• Diffusion MRI
• Artificial Intelligence in Medicine

Background:

• Intravoxel incoherent motion (IVIM) diffusion-weighted imaging is crucial for monitoring treatment response.
• Traditional IVIM fitting methods are noise-sensitive and lack spatial correlation, leading to unstable parameter estimation.
• Existing deep learning methods have limitations due to local receptive fields.

Purpose of the Study:

• To develop a robust two-stage IVIM parameter estimation framework using Implicit Neural Representation (IVIM-INR).
• To enhance spatial context modeling and overcome limitations of traditional and existing deep learning approaches for IVIM analysis.

Main Methods:

• The IVIM-INR framework utilizes coordinate encoding for global spatial perception.
• It incorporates local 3D patch information from multi-b-value images for improved spatial context.
• A two-stage approach involves signal denoising followed by accurate IVIM parameter fitting.

Main Results:

• IVIM-INR demonstrated significant advantages on brain phantoms, AAPM breast data, and clinical Glioblastoma (GBM) data.
• Reduced normalized mean absolute errors (NMAEs) for Dp, Dt, and Fp in simulated brain tumor regions compared to other methods.
• Achieved a 58% reduction in Fp error compared to ConvNet in breast tumor tissues and high ICC values for Dt in clinical data.

Conclusions:

• IVIM-INR combines the continuous modeling of INR with spatial-aware features for improved IVIM analysis.
• The method effectively overcomes limitations of traditional techniques in noisy conditions.
• Provides a more reliable tool for quantitative IVIM analysis in clinical settings.