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

BACKGROUND

Intravoxel incoherent motion (IVIM) imaging, a diffusion magnetic resonance imaging technique, is commonly used to quantify tissue perfusion and diffusion. Traditional pixel-by-pixel fitting methods, however, often suffer from high noise, causing unreliable parameter estimates.

PURPOSE AND

METHODS

To address this issue, a novel unsupervised learning-based framework combining a one-dimensional (1D) convolutional neural network (CNN) with a bidirectional long short-term memory (BiLSTM) network and a multi-head attention mechanism (MHAM) was proposed. Several techniques were proposed to reduce the effect of random weight initialization, noisy input data, and overfitting/underfitting on the estimation of IVIM parameters. The performance of the proposed method was evaluated using both simulated and experimental data, and the results were compared with those obtained using the deep neural network (DNN) method and the Bayesian-Markov random fields (MRF) method.

RESULTS

Simulation results showed that the proposed method achieved lower root mean square error values than the other two methods, indicating more reliable IVIM parameter estimates. The only exception was at a signal-to-noise ratio of 100, where it performed similarly to the Bayesian-MRF method. For the abdominal datasets, the proposed method yielded IVIM parameter estimates that closely matched the literature-reported values and avoided the overestimation of pseudo-diffusion coefficients (D*) observed in the other two methods. For the brain dataset, the perfusion fractions and diffusion coefficients obtained from all three methods were consistent with the literature-reported ranges; however, only the DNN method tended to overestimate D*.

CONCLUSIONS

These findings suggest that the proposed CNN-BiLSTM-MHAM model is a promising approach for IVIM parameter estimation.