Unsupervised 1D CNN -bidirectional long short-term memory model with multi-head attention for generating intravoxel...

Area of Science:

• Medical Imaging
• Machine Learning

Background:

• Intravoxel incoherent motion (IVIM) imaging is a diffusion MRI technique used for quantifying tissue perfusion and diffusion.
• Traditional pixel-by-pixel fitting methods for IVIM analysis are susceptible to noise, leading to unreliable parameter estimates.

Purpose of the Study:

• To develop a novel unsupervised learning framework to enhance the accuracy and reliability of IVIM parameter estimation.
• The framework aims to mitigate issues related to noisy data, random initialization, and model overfitting/underfitting.

Main Methods:

• A hybrid deep learning architecture integrating a 1D convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM) network, and multi-head attention mechanism (MHAM) was proposed.
• The model's performance was validated using both simulated and experimental abdominal and brain MRI datasets.
• Comparative analysis was conducted against deep neural network (DNN) and Bayesian-Markov random fields (MRF) methods.

Main Results:

• The proposed CNN-BiLSTM-MHAM model demonstrated lower root mean square error in simulations compared to DNN and MRF methods, except at high signal-to-noise ratios.
• For abdominal data, the model produced IVIM parameters consistent with literature values and avoided pseudo-diffusion coefficient overestimation.
• Brain dataset analysis showed consistent perfusion and diffusion coefficients, with the proposed model avoiding D* overestimation seen with the DNN method.

Conclusions:

• The developed CNN-BiLSTM-MHAM model presents a robust and promising approach for accurate intravoxel incoherent motion parameter estimation.
• This deep learning framework offers improved reliability for diffusion and perfusion quantification in MRI applications.