IVIM parameters mapping with artificial neural network based on mean deviation prior.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Intravoxel Incoherent Motion (IVIM) imaging offers promising biomarkers for disease management.
• Supervised learning methods for IVIM parameter estimation suffer from performance degradation due to the out-of-distribution (OOD) problem.
• A gap exists between simulated and real-world IVIM datasets, hindering clinical applications.

Purpose of the Study:

• To propose IterANN, a novel learning framework to overcome the OOD problem in IVIM parameter estimation.
• To enhance the accuracy and stability of IVIM parameter estimation using a mean deviation prior (MDP).
• To improve the clinical applicability of IVIM imaging for disease diagnosis.

Main Methods:

• IterANN utilizes a simple artificial neural network (ANN) with an input layer for multi-b-value signals and an output layer for three IVIM parameters.
• The framework incorporates a mean deviation prior (MDP) to iteratively update the distribution of simulated training data.
• The training data distribution is adjusted to align its mean with predicted real data values, enhancing correlation.

Main Results:

• IterANN achieved the lowest residual error in IVIM parameters on simulation datasets, particularly at low SNR.
• The method demonstrated a significant reduction in residual error for key IVIM parameters compared to suboptimal approaches.
• On real datasets, IterANN yielded the highest parameter contrast-to-noise ratio (PCNR) and superior stability with lower coefficients of variation (CV).

Conclusions:

• Updating training data distribution using MDP effectively resolves the OOD problem in IVIM analysis.
• IterANN significantly improves the accuracy and stability of estimated IVIM parameters.
• The enhanced performance of IterANN increases the potential of IVIM imaging for disease diagnosis.