SynQ-Breast: a flexible deep learning framework for quantitative MRI of breast tumors via synthetic data with IVIM-MRI validation.
Authors
Affiliations (5)
Affiliations (5)
- Shenzhen Research Institute of Xiamen University, Shenzhen, China.
- Department of Electronic Science, Fujian Provincial Key Laboratory of Plasma and Magnetic Resonance, Xiamen University, Xiamen, China.
- Department of Computer Science, Medical Imaging Technology Section, Jiangxi University of Chinese Medicine, Nanchang, China.
- Department of Magnetic Resonance Imaging, The First Affiliated Hospital of Zhengzhou University, Zhengzhou, China.
- School of Ocean Information Engineering, Fujian Provincial Key Laboratory of Maritime Communication and Intelligent Electronic Systems, Jimei University, Xiamen, China.
Abstract
Breast tissue exhibits inherently low signal-to-noise ratio (SNR) on diffusion-weighted imaging, which compromises the accuracy and stability of conventional model fitting. This study aimed to develop a flexible synthetic-data-driven deep learning framework (SynQ-Breast) for robust quantitative magnetic resonance imaging (qMRI) parameter estimation of breast tumors, addressing the challenges of low SNR and the scarcity of realistic training data. The supervised SynQ-Breast framework incorporates a training data synthesis module, where synthetic data are generated based on complex parameter distributions and spatial tissue texture. The synthetic data were modulated with realistic Rician noise to match clinical acquisitions, enabling the model to learn an anti-noise mapping to qMRI parameters. A U-Net incorporating spatial smoothness constraint was trained to produce qMRI parametric maps. Validation was performed on intravoxel incoherent motion MRI (IVIM-MRI) data from 49 breast tumor patients. Diagnostic performance was assessed using first-order histogram metrics from tumor regions, supported by statistical analyses (Mann-Whitney U test, independent <i>t</i>-test, Pearson correlation analysis, receiver operating characteristic curve analysis, and binary logistic regression analysis). SynQ-Breast (30 ms per slice) achieved a three-order-of-magnitude acceleration in computational time, which was statistically significant (P<0.0001) compared to conventional nonlinear least squares (NLLS; 44 s per slice). The framework also yielded better lesion detectability in breast tissue in term of contrast-to-noise (CNR) [SynQ-Breast (mean ± standard deviation): 1.67±1.83, 1.76±1.58, 1.66±2.17, NLLS (mean ± standard deviation): 1.11±0.85, 0.47±0.37, 0.63±0.40 for diffusion coefficient (<i>D</i>), perfusion fraction (<i>f</i>), pseudo-diffusion coefficient (<i>D</i>*) respectively]. For tumor differentiation, SynQ-Breast achieved higher area under the curve (AUC) values than NLLS for single metric (best AUC: 0.843 <i>vs.</i> 0.808) and combined metric (AUC: 0.924 <i>vs.</i> 0.879). The SynQ-Breast framework provides a robust and efficient solution for breast qMRI, overcoming data scarcity and noise sensitivity. The performance of SynQ-Breast in IVIM-based breast tumor differentiation suggests its potential clinical value.