Room: Exhibit Hall | Forum 8
Purpose: Conventional approaches to estimate parameters in the intravoxel incoherent motion (IVIM) model for diffusion-weighted magnetic resonance imaging (DW-MRI) are to apply fitting algorithms to the multi b-value DWI, which could be sensitive to noise in the signals and fitting options. This study aimed to develop a new parameter estimation method based on machine learning to achieve better accuracy and robustness in the DW-MRI analysis.
Methods: A convolutional neural network (CNN) model is built and trained using simulated DWI signal attenuation curves. The proposed CNN model has three convolutional layers followed by two fully connected layers. Input to the network has intensities from 11 b-values and the output layer has three nodes, each for the pseudo diffusion coefficient D*, true diffusion coefficient D, and perfusion fraction f in the IVIM model. A new simulation dataset with added Gaussian noise is used for evaluation of the CNN model. Mean absolute error (MAE) and mean squared error (MSE) are computed for each parameter for the performance evaluation. A curve fitting method based on the Nelder-Mead method (a.k.a, simplex) is used to compare the performance of the proposed CNN model.
Results: The CNN model produced lower MAE and MSE for parameters D and f while yielded slightly higher error rates for D* than the simplex fit. Also, the model showed consistent performance on a dataset containing parameter values from extended D* range, which suggests that the model successfully found a mapping function to estimate the parameters in the IVIM method.
Conclusion: The proposed CNN model showed promising initial results for the parameter estimation in DW-MRI data analysis. Further investigation is needed to (i) find a better model architecture to improve the performance, (ii) evaluate the performance of the model on actual patient data, and (iii) apply the technique to other analysis models.
Funding Support, Disclosures, and Conflict of Interest: Other than reporting the funding support from NIH UO1 CA183848, we don't have anything to disclose.