Parameter Estimation for the Intravoxel Incoherent Motion-diffusional Kurtosis Imaging Model Using Synthetic Q-space Learning Toward Breast Tumor Diagnosis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Medical Image Computation, Tohoku University Graduate School of Medicine, Sendai, Miyagi, Japan.
- Department of Diagnostic Radiology, Tohoku University Hospital, Sendai, Miyagi, Japan.
- Department of Radiology, Akita University Graduate School of Medicine, Akita, Akita, Japan.
- Division of Image Statistics, Tohoku Medical Megabank Organization, Tohoku University, Sendai, Miyagi, Japan.
- Department of Diagnostic Radiology, Tohoku University Graduate School of Medicine, Sendai, Miyagi, Japan.
Abstract
Synthetic Q-space learning (synQSL), which uses only synthetic data in the training of regressors, has demonstrated promising results in parameter estimation for brain diffusion MRI (dMRI). This study aimed to evaluate the performance of synQSL in breast dMRI for intravoxel incoherent motion-diffusional kurtosis imaging (IVIM-DKI) parameter estimation, with comparing several types of regressors and conventional fitting by nonlinear least squares fitting (NL-LSF). In synthesizing data for synQSL, IVIM-DKI parameters were sampled from uniform distributions and substituted into the signal model along with b-values to generate diffusion-weighted imaging (DWI) signals. In addition, Rician noise was mixed to the signals. We prepared datasets of 10<sup>5</sup> and 10<sup>6</sup> samples, and trained multi-layer perceptron (MLP), Kolmogorov-Arnold networks (KAN), and random forest (RF) regressors for synQSL. The performance of the parameter estimation methods including NL-LSF was evaluated using a digital phantom including various value combinations of IVIM-DKI parameters and clinical data of 67 cases (13 benign and 54 malignant lesions) through quantitative analysis and visual assessment. In the digital phantom, the regressors of synQSL achieved significantly lower root mean square error (RMSE) for f, D<sup>∗</sup>, and D than NL-LSF (P < 0.05) with less variation among the parameter sets. In clinical datasets, synQSL not only improved the visual quality of parameter maps but also showed significant differences between benign and malignant lesions in parameters while the NL-LSF failed. Furthermore, the estimation times for all synQSL regressors were substantially shorter than that of NL-LSF. Among the synQSL regressors, MLP showed superior properties including computational cost. synQSL demonstrated superior parameter estimation performance compared to NL-LSF in breast IVIM-DKI analysis. In this study, MLP was considered the most suitable regressor for synQSL among those we examined, based on the balance between estimation accuracy and computational costs.