SynQ-Breast: a flexible deep learning framework for quantitative MRI of breast tumors via synthetic data with IVIM-MRI validation

Introduction

Breast cancer has emerged as the second most frequently diagnosed cancer, posing a major public health challenge for women (1), and early and accurate differentiation between benign and malignant breast lesions is crucial for optimizing treatment strategies and improving patient prognosis (2). While percutaneous core needle biopsy remains the diagnostic gold standard, its clinical utility can be limited by potential sampling error and technical challenges in complex cases, possibly leading to instances of both missed diagnosis and misdiagnosis (3). In this regard, magnetic resonance imaging (MRI)-based methods for breast cancer differentiation are desired for its noninvasiveness and high imaging freedom. Existing literature has reported the clinical value of apparent diffusion coefficient (4), diffusion kurtosis imaging (5), dynamic contrast-enhanced MRI (DCE-MRI) derived parameters (6) in differentiating breast cancers for their pathological associations with microstructural and functional changes of abnormal tissues. Moreover, it has also been reported that these parameters, when used with multiparametric models (7,8), or with machine learning and radiomics (9,10), improve diagnostic accuracy.

However, the diagnostic potential of these imaging methods is heavily dependent on the accuracy and precision of the estimated parameters, which may be collectively compromised by the sensitivity of nonlinear least squares (NLLS; the most widely used approach for MRI parameter estimation) to noise and the inherent low signal-to-noise ratio (SNR) of diffusion-weighted breast, leading to suboptimal parametric map quality and poor reproducibility (11). Of note, the breast is a heterogeneous mix of fibroglandular and fatty tissues with substantially different proton densities, and a higher fat fraction directly results in lower voxel wise proton density and thus lower achievable SNR (12). This issue can be worse when fat suppression pulses (13) or high spatial-resolution protocols (14) are applied. Meanwhile, B₁ field inhomogeneity (15,16), particularly in the retro areolar and chest wall regions or at high-field strengths, along with patient positioning and gravitational tissue shift can give rise to notable SNR variation (17).

Recently, deep learning has been applied to estimate diffusion parameters (18-26), but still faces challenges such as high noise vulnerability [e.g., unsupervised methods (20,22,23), which have been proven to be ill-suited to low-SNR breast MRI (27)] or requiring abundant high-quality labeled data for model training [e.g., supervised learning (28)]. To improve the clinical practicability of supervised learning, synthetic data have been used (19,24,25,29-32). However, some voxel-wise deep learning methods (19,29,30) for diffusion parameter estimation report noisy parameter maps and high sensitivity to noise, especially in low-SNR tissues. Furthermore, some synthetic data generation strategies rely exclusively on uniform prior distributions, which can lead to unrealistic parameter distributions (28). To address these issues, several studies (24,25,32) have demonstrated that incorporating spatial smoothness priors can substantially improve parameter estimation robustness.

Herein, we introduce a flexible synthetic-data-driven deep learning framework, termed SynQ-Breast, to implement noise-insensitive quantitative magnetic resonance imaging (qMRI) parameter estimation of breast tumors, while evading the clinical infeasibility of acquiring paired data samples for model training and the unrealistic distributional priors of data. Technically, SynQ-Breast incorporates more complex parameter distribution and spatial texture priors, designed to better simulate the intricate appearance of both normal tissue and breast lesions. The inputs of synthetic training samples were modulated with Rician noise at levels representative of clinical breast MRI acquisitions, while the labels remained noise-free. This strategy enables the model to learn an anti-noise mapping, allowing it to accurately estimate quantitative parameters from clinical data without noise interference. While the SynQ-Breast framework is designed to be diffusion MRI (dMRI)-model-agnostic, in this work, we apply and validate it specifically for the intravoxel incoherent motion (IVIM) model parameter estimation using a clinical cohort of 49 breast tumor patients. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0275/rc).

Methods

IVIM biexponential model

The IVIM biexponential model facilitates the independent quantification of two distinct physiological processes: (I) molecular diffusivity, represented by the true diffusion coefficient (D), and (II) microcapillary perfusion, characterized by both flow-related pseudo-diffusion coefficient (D*) and perfusion fraction (f), thereby providing a more comprehensive assessment of tissue microstructure and vascular properties. Since Sigmund et al.’s pioneering application of IVIM-MRI to breast cancer in 2011 (33), this technique has gained increasing recognition in breast imaging. Current clinical applications of IVIM-MRI involve differentiation of benign and malignant lesions through quantitative perfusion and diffusion analysis, identification of molecular prognostic markers for personalized treatment planning, and prediction and monitoring of therapeutic response to guide clinical decision-making (34-36). The IVIM bi-exponential model is as follows:

S(b)=S0[(1−f)e−bD+fe−bD*][1]

where b is the diffusion sensitivity factor, S₀ is the signal intensity acquired without applying any diffusion-sensitizing gradients. It should be noted that the IVIM bi-exponential model is a simplified model and does not account for compartment-specific relaxation effects, which has been recognized as a limitation in previous studies (37,38). In particular, the tissue and vascular/perfusion-related compartments may have different T₂ relaxation times. To address this issue, a more complete IVIM model incorporating echo time (TE)-dependent relaxation terms for different compartments has been proposed (39). Nevertheless, implementation of such relaxation-corrected IVIM model typically increases scan time, thereby limiting its routine clinical applicability.

Patients

This prospective study enrolled patients with breast tumors from The First Affiliated Hospital of Zhengzhou University between December 2015 and December 2017. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of The First Affiliated Hospital of Zhengzhou University (No. 2019-KY-231). Informed consent was obtained from all individual participants included in the study. Patient demographics and clinical characteristics are summarized in Table 1. The patient inclusion process is detailed in Figure 1. All patients in this study underwent diagnostic breast MRI based on clinical indications. Pathological confirmation (surgical resection or core-needle biopsy) was obtained within two weeks after the MRI scan for every case. 16 patients were excluded because of poor image quality, such as low SNR and artifacts. No high-risk breast lesions were identified. 49 lesions were detected in 46 patients. There were 27 benign lesions in 24 patients, including 17 fibroadenomas, 4 adenosis, 3 intraductal papillomas, and 3 inflammations of breast. The remaining 22 lesions were malignant in 22 patients, which included 18 non-specific invasive ductal carcinomas and 4 intraductal carcinomas. For the 12-year-old patient, she was histopathologically confirmed to have intraductal papilloma.

Figure 1 Flow chart of patient inclusion. MRI, magnetic resonance imaging.

MRI protocols

Breast MRI examinations were conducted on a 3.0 T MRI scanner (GE Discovery 750, GE Healthcare, Milwaukee, WI, USA) equipped with an 8-channel dedicated breast coil. All patients were positioned prone during imaging. Axial T₁-weighted images were acquired with the following parameters: repetition time (TR)/TE =640/7.6 ms; field of view (FOV) =320×320 mm²; matrix =320×320; slice thickness/gap =4/1 mm; with spectral fat saturation. T₂-weighted imaging was performed using: TR/TE =2,587/85 ms; FOV =320×320 mm²; matrix =320×192; slice thickness/gap =4/1 mm. Diffusion-weighted imaging (DWI) employed 14 b-values (0, 20, 50, 100, 150, 200, 400, 800, 1,200, 1,600, 2,000, 2,500, 3,000, 4,000 s/mm²) with corresponding number of excitations (NEX) of 1, 1, 1, 1, 1, 2, 2, 2, 4, 4, 6, 6, 8, 10; TR/TE =3,600/76 ms; FOV =320×320 mm²; matrix =128×192; slice thickness/gap =4/1 mm. Dynamic contrast-enhanced (DCE) MRI was performed using an axial VIBRANT sequence with TR/TE =3.9/1.7 ms; FOV =360×360 mm²; matrix =320×320; slice thickness/gap =1.4/1.0 mm. Contrast agent (Gd-DTPA) was administered intravenously at 0.1 mmol/kg body weight using a power injector at 2.0 mL/s, followed by a 20 mL saline flush. Dynamic acquisition started 20 s after contrast injection, consisting of five sequential phases, each with a duration of 59 s.

Training data preparation

Figure 2A illustrates the overview of SynQ-Breast for IVIM-MRI parameter mapping and clinical value evaluation, and Figure 2B show the neural network architecture employed in SynQ-Breast. Our synthetic data generation strategy comprised three key phases. First, we constructed basic IVIM parametric maps by randomly generating geometric primitives (including circles, squares, triangles, and rings) with stochastic locations and sizes within a defined mask region. The geometric primitives are continuously placed until the mask region is filled by more than 80%. These primitives were assigned IVIM parameters (D: 0.00004–0.0015 mm²/s, D*: 0.0015–0.15 mm²/s, f: 0.003–0.303) sampled from hypothetical uniform distributions. Second, to bridge the gap between simple geometric primitives and real anatomical structures, we integrated relatively complex texture derived from an optical image database including 2,201 different optical images. Although other medical images may provide more realistic texture, optical images were used because they were available in large quantities at very low cost. With a sufficiently large and diverse dataset, optical images can effectively cover a wide range of texture variations in medical images. Specifically, the optical images were converted to grayscale and processed with Gaussian filtering. The resulting spatial textures were then added to the quantitative primitives via weighted summation. This texture modulation, as shown in Figure 3, introduced structure variations, effectively transforming simple distribution of IVIM parametric maps into complex spatial heterogeneity observed in clinical breast MRI. Finally, we synthesized multi-b-value diffusion-weighted (DW) images through the IVIM model using the synthetic IVIM parametric maps as templates. Rician noise was added to all synthesized DW images to match the noise characteristics and level (SNR =17 dB) of clinical acquisition. The resulting data contained paired samples of noise-modulated DW images and their corresponding noise-free ground-truth parametric maps for supervised network training.

Figure 2 Method overview and neural network architecture. (A) Overview of SynQ-Breast for IVIM-MRI parameter estimation and clinical value evaluation. (B) Neural network architecture employed in SynQ-Breast. BN, batch normalization; D, diffusion coefficient; D*, pseudo-diffusion coefficient; DW, diffusion-weighted; f, perfusion fraction; IVIM, intravoxel incoherent motion; MRI, magnetic resonance imaging; ReLU, rectified linear unit.

Figure 3 Example of texture modulation. λ is the weight coefficient randomly sampling between 0.7 and 1.

Neural network training and IVIM parameter estimation

To establish an effective nonlinear mapping between the source multi-b-value DW images and the targeted IVIM-MRI parametric maps, a 5-level U-Net with skip connections (Figure 2B) was implemented. This network design, featuring a characteristic encoder-decoder structure, has been extensively validated in various medical image tasks, particularly in image segmentation and cross-modality transformation (40). The training process utilized a synthetic dataset comprising 3,300 samples, of which 3,000 were used for training, and the remaining 300 were used for validation. The input to the network consisted of the synthetic noise-modulated DW images, while the training label was the corresponding synthetic noise-free IVIM-MRI parametric maps. Both the input and output images had a matrix size of 128×128 during training. During testing, the input was clinical DW images with a reconstructed matrix size of 128×192 without any reshaping or resizing, as the fully convolutional architecture of the U-Net supported the trained model to directly infer on different matrix size without affecting model performance. The optimization process employed mean square error (MSE) as the loss function, which quantifies the pixel-wise differences between the network predicted parametric maps and the corresponding ground truth parametric maps. The Adam optimizer with momentum parameters β₁ =0.9 and β₂ =0.999 was used to update the network parameters during training. To enhance the learning efficiency and prevent overfitting, several strategies were incorporated: the batch size was maintained at 8, and the learning rate was initially 10⁻⁴ and gradually decreasing by 20% after every 30,000 iterations until convergence of the neural network.

Histogram analysis and clinical value evaluation

Histogram analysis and clinical value evaluation were finished based on the regions of interest (ROIs) of parametric maps. ROI placement was performed manually on DW images with b value of 800 s/mm², with careful reference to corresponding DCE images to exclude cystic or necrotic regions. ROIs were manually delineated on one slice showing the largest lesional cross-sectional area, then precisely transferred to corresponding parametric maps (D, D*, and f) for quantitative analysis. ROI placement was performed independently by two experienced medical imaging researchers (with 3 and 5 years of experience in breast MRI, respectively). In cases of disagreement, consensus was reached through joint review and discussion. To enable comparative assessment of lesion characterization capability between SynQ-Breast and NLLS methods, IVIM-MRI parametric mapping was performed for both the entire breast parenchyma and specific lesion areas, and the diagnostic performance of IVIM-MRI parameters in differentiating benign from malignant lesions were quantitatively evaluated within the defined ROIs.

Comparative method

The conventional NLLS algorithm was used to fit IVIM-MRI parameters using in-house built code based on MATLAB 2018a (Mathworks, Natick, MA, USA). Following the established methodology by Sigmund et al. (33), a segmented fitting approach was employed to enhance the reliability and validity of parameter estimation. This method is based on the physiological assumption that pseudo-diffusion effects become negligible at high b-values (b >200 s/mm²), allowing for separate estimation of diffusion and perfusion parameters. Specifically, the diffusion coefficient (D) was calculated using the simplified mono-exponential model applied to high b-value data:

S=Sinte−bD[2]

where S_int is the zero-intercept value derived from mono-exponential fitting of the high b-value data. Subsequently, the perfusion fraction (f) was estimated using the following formula:

f=S0−SintS0[3]

Finally, the estimated parameters D and f were incorporated into the bi-exponential IVIM-MRI model to estimate D* through NLLS fitting. This sequential estimation approach ensures numerical stability while maintaining physiological consistency in the parameter estimation process.

The contrast-to-noise ratio (CNR) between the lesion and the normal region surrounding the lesion was used to quantify the quality of estimated parametric maps in the tumor region. The formula is as follows (41):

CNR=mean(θlesion)−mean(θnormal)std(θnormal)[4]

where θ_lesion and θ_normal represent the estimated parameters of the lesion ROI and the 5-pixel-wide annular normal region surrounding the lesion, respectively. And mean and std indicate calculating the average and standard deviation (SD).

Statistical analysis

Statistical analyses were conducted using SPSS 25.0 (SPSS, Chicago, IL, USA) and MedCalc 20.0.22 (MedCalc Software, Mariakerke, Belgium). Quantitative analysis of ROI characteristics was conducted through histogram analysis, extracting key metrics including mean value, extreme value, and heterogeneity indices. All variables underwent statistical evaluation, including Shapiro-Wilk normality test and the Levene homogeneity of variance test. Normally distributed data were expressed as mean ± SD and compared using independent samples t-test, while non-normally distributed data were presented as median (interquartile range) and analyzed using Mann-Whitney U test. The correlation between the IVIM-MRI parameters estimated from SynQ-Breast and the NLLS method was assessed using Pearson correlation analysis. Diagnostic performance was evaluated through receiver operating characteristic (ROC) curve analysis, with the area under the curve (AUC), sensitivity, and specificity calculated based on the Youden index maximization criterion (Youden index = sensitivity + specificity − 1). The binary logistic regression analysis was employed to combine IVIM-MRI parameters and histogram indices for distinguishing between benign and malignant lesions. All statistical tests were conducted at a significance level of α=0.05.

Results

Comparison of IVIM-MRI parametric map quality

Figure 4 presents comparative results of IVIM-MRI parametric maps estimated from both methods for 4 patients with breast tumor. The estimated D parametric maps showed good visual consistency between SynQ-Breast and conventional NLLS, whereas the D* and f maps demonstrated more noticeable visual differences between the two methods. The network derived D maps exhibit superior image quality, characterized by enhanced lesion boundary definition and reduced noise level compared to the NLLS method. Notably, the D* maps estimated from SynQ-Breast reveal elevated parameter values at lesion peripheries, providing improved margin delineation. A critical limitation of the NLLS method is evident in the f maps, where numerous voxels reach the upper threshold of the predefined perfusion fraction range, resulting in poor lesion to breast tissue contrast and compromised diagnostic utility. In contrast, SynQ-Breast yields f maps with clearly visible lesion margin and contrast resolution comparable to D maps. Across all breast cancer cases, SynQ-Breast achieved higher lesion-to-normal CNR (mean ± SD) than NLLS for all three IVIM parameters: D (1.67±1.83 vs. 1.11±0.85), f (1.76±1.58 vs. 0.47±0.37), and D* (1.66±2.17 vs. 0.63±0.40). These results indicate that SynQ-Breast provides substantially better lesion conspicuity and tissue contrast than conventional NLLS fitting. In addition, the computational efficiency of SynQ-Breast exhibits a substantial improvement, shortening the parameter reconstruction duration from approximately 44 s per slice with NLLS method (range: 30–60 s) to about 30 ms per slice (range: 24–35 ms), representing a three-order-of-magnitude acceleration in processing speed (P<0.0001).

Figure 4 IVIM-MRI parametric maps of 4 patients with breast tumor estimated from NLLS and SynQ-Breast. D, diffusion coefficient; D*, pseudo-diffusion coefficient; f, perfusion fraction; IVIM, intravoxel incoherent motion; MRI, magnetic resonance imaging; NLLS, nonlinear least squares.

Comparison of IVIM-MRI quantitative parameters

Table 2 shows a comparative analysis of IVIM-MRI histogram metrics between benign and malignant breast lesions derived from both methods (only histogram metrics with significant differences between benign and malignant breast lesions were given). It is obvious that among the parameters estimated from NLLS, D (skewness, kurtosis), D* (skewness), and f (variance) exhibit significantly lower values in benign lesions, while D (minimum, mean, median) and f (minimum, mean) show higher values compared to malignant lesions. The SynQ-Breast method reveals distinct patterns, D* (maximum, mean, median, variance) are significantly reduced in benign lesions, whereas D (mean, median) and f (maximum, minimum, mean, median) are significantly higher in benign breast lesions than in malignant lesions.

Correlation analysis between the two methods, focusing on mean and median values of each parametric map (Table 3), reveals strong concordance for D values, with correlation coefficients of 0.893 (P<0.0001) for D-mean and 0.897 (P<0.0001) for D-median. In contrast, D* and f show no significant correlations between the two methods.

Figure 5 illustrates the comparative scatter plots of mean and median IVIM-MRI parameters across all cases. While D and D* distributions demonstrate general alignment between both methods, the NLLS method exhibits notable outliers, indicating estimation instability. SynQ-Breast shows superior concentration with fewer outliers, reflecting enhanced estimation stability. Additionally, the scatter plots highlight a critical limitation of NLLS method in f parameter estimation, where a substantial proportion of voxels cluster at the upper boundary of the predefined range, potentially compromising diagnostic accuracy.

Figure 5 Scatter plot analysis comparing IVIM-MRI parameters estimated from NLLS and SynQ-Breast methods. ns, P≥0.05; **, P<0.01. D, diffusion coefficient; D*, pseudo-diffusion coefficient; f, perfusion fraction; IVIM, intravoxel incoherent motion; MRI, magnetic resonance imaging; NLLS, nonlinear least squares; ns, no significant difference.

Comparison of diagnostic efficiency

Figure 6 shows the ROC curves comparing the diagnostic performance of both methods in differentiating benign and malignant breast lesions, with corresponding quantitative metrics summarized in Table 4. Analysis of the NLLS derived parameters reveals that D and f histogram metrics demonstrate superior diagnostic efficacy compared to D*, consistent with previous findings (36,42,43). In contrast, SynQ-Breast results show enhanced diagnostic performance, particularly in D* map. SynQ-Breast achieves optimal diagnostic performance with D*-median, demonstrating an AUC of 0.843 and classification accuracy of 81.6%. Comparatively, the NLLS method’s best-performance parameter, D-median, yields an AUC of 0.808 and accuracy of 79.6%. Several single-parameter statistics (e.g., D-mean and f-mean) showed marginally higher classification performance in NLLS. This is possibly due to the relatively higher SNR in some lesions and incomplete separation of D* from D and f in NLLS. SynQ-Breast yielded more robust parametric maps under the inherently low-SNR conditions of breast tissue and better diagnostic accuracy when more parameters were combined. The NLLS method reaches maximal performance (AUC =0.879, accuracy =85.7%) through combination of D-skewness, D*-skewness, and f-variance, while SynQ-Breast achieves superior diagnostic efficacy (AUC =0.924, accuracy =89.8%) by integrating D*-maximum, D*-median, D*-variance and f-median. These results demonstrate that SynQ-Breast not only improves individual parameter performance but also enables more effective parameter combination for enhanced diagnostic accuracy in breast lesion characterization.

Figure 6 ROC comparative analysis of combined-parameter diagnostic performance between NLLS and SynQ-Breast. AUC, area under the curve; NLLS, nonlinear least squares; ROC, receiver operating characteristic.

Discussion

Conventional voxel-wise IVIM-MRI parameter fitting methods face several inherent limitations, including prolonged computational time, loss of image structure information and susceptibility to noise. These inherent constraints have largely restricted the application of IVIM-MRI parametric maps in breast lesion detection, with most studies focusing on differentiating between benign and malignant lesion. The implementation of deep neural network in SynQ-Breast effectively addresses these challenges, yielding parametric maps with enhanced structural clarity and reduced background noise.

Recently, the voxel-based deep learning method for IVIM-MRI has attracted interest. To further validate the superiority of the SynQ-Breast we conducted an image quality comparison between SynQ-Breast and a representative voxel-based deep learning approach (IVIM-NET) (22) trained on synthetic data, in addition to conventional NLLS fitting. SynQ-Breast achieved the highest CNR (mean ± SD) across all three IVIM parameters (D: 1.67±1.83; f: 1.76±1.58; D*: 1.66±2.17), substantially ouperforming both NLLS (D: 1.11±0.85; f: 0.47±0.37; D*: 0.63±0.40) and IVIM-NET (D: 1.29±0.97; f: 0.36±0.24; D*: 0.33±0.22). Although IVIM-NET slightly outperformed NLLS in the D parameter, it showed even poorer performance than NLLS for the f and D* parameters. These results suggest that the voxel-based deep learning method provides limited advantages in challenging low-SNR regions such as breast tissue. It has been noted that IVIM-NET exhibited clear biases, particularly for the D* parameter in low-SNR regions, and showed instability across IVIM parameters (44). In contrast, the image-based SynQ-Breast, which utilizes spatial information, yields markedly better parametric map quality. This observation is consistent with previous studies showing that spatial-context-aware methods are generally more robust to noise than purely voxel-based approaches (23,26,45). The SD of CNR values across different patients reflects the biological heterogeneity of the cohort. A wider CNR range of SynQ-Breast indicates that it provides a broader dynamic range of tissue contrast across different lesion types, which is also consistent with its superior diagnostic performance. SynQ-Breast method achieved its best performance with D*-median (AUC =0.843). In contrast, NLLS performed best with D-median (AUC =0.808), while IVIM-NET yielded all AUCs below 0.75 (best: D-skewness =0.722). These indicate that SynQ-Breast outperforms the other two methods in perfusion-related parameter. In addition, SynQ-Breast demonstrated the best combined-parameter diagnostic performance (AUC =0.924), outperforming NLLS (AUC =0.879) and IVIM-NET (AUC 0.786 by integrating D-median, D*-variance and f-variance).

Comparative analysis reveals strong consistency in D parametric maps between SynQ-Breast and NLLS methods, with both effectively delineating lesions across breast tissue. The results are consistent with and further substantiate previous studies that have confirmed the reliability and robustness of parameter D estimation (42,46,47). Notably, SynQ-Breast shows remarkable improvement in D* map quality, particularly at lesion peripheries where increased parameter values enhance margin visibility. This phenomenon may be attributed to increased microvascular density and blood flow velocity at lesion edges, as suggested by existing physiological models (48). Another important finding is observed in f maps, where SynQ-Breast overcomes the limitations of NLLS fitting, particularly its susceptibility to extreme values and low-quality voxel data (42). The resulting f maps exhibit high image quality, comparable to D maps in structural definition and diagnostic utility. The improvements in parametric map quality, characterized by enhanced visual smoothness and anatomical fidelity, represent a substantial step toward increased clinical feasibility of IVIM-MRI technology for breast imaging.

The SynQ-Breast framework represents a significant advancement in IVIM-MRI parameter estimation. The modular design of SynQ-Breast allows straightforward adaptation to other quantitative models through simple substitution of the corresponding physical model in the data generation pipeline (Figure 2A). This model-agnostic characteristic positions SynQ-Breast as a versatile platform for various qMRI applications. Successful implementation requires careful consideration of two fundamental aspects: physiologically appropriate parameter boundaries and representative noise characteristics that match clinical reality.

Our investigation reveals several crucial insights regarding parameter selection for IVIM modeling. First, maintaining the physiological relationship between diffusion and perfusion parameters is essential. Specifically, the minimum D* value should meet or exceed the maximum D value to reflect the intrinsic property that pseudo-diffusion typically exceeds pure molecular diffusion. Second, setting the minimum f value below the lowest expected perfusion fraction in breast tissue ensures adequate sampling of the hypo-perfused tissue spectrum. These empirically derived guidelines underscore the importance of aligning mathematical parameterization with underlying tissue physiology, a principle that may be extended to other qMRI techniques.

The challenge of varying SNRs across clinical settings warrants particular attention. Differences in scanner hardware, acquisition protocols, and institutional practices create substantial SNR variability that directly impacts parameter estimation accuracy. To address this, we implement the bisection method to determine optimal noise levels for synthetic data generation, with parametric map quality serving as the primary feedback. This process reveals distinct characteristic: underestimation of clinical SNR would produce oversmoothed parameter maps with loss of anatomical detail, while overestimation would result in noise amplification and compromised quantitative accuracy. Notably, despite the absence of precise SNR quantification in our clinical data, SynQ-Breast demonstrates robust performance, highlighting its inherent adaptability to diverse noise environments.

The generalization capability of SynQ-Breast stems from its innovative distribution-agnostic data generation strategy. By constructing complex parameter distributions through aggregation of random geometrical primitives rather than replicating specific anatomical patterns, our method avoids the common drawback of overfitting to particular dataset characteristics. This approach creates rich and varied training data that emphasizes fundamental tissue property relationships over specific anatomical appearances. Its practical implementation requires adjustment of only three manageable factors: clinically relevant parameter value ranges, appropriate SNR levels, and specific acquisition parameters. This streamlined customization process facilitates the adaptation of SynQ-Breast to new clinical scenarios while maintaining estimation reliability, although we acknowledge the inherent trade-off between broad applicability and optimal performance in specific contexts.

From the perspective of clinical translation, SynQ-Breast offers substantial practical advantages. The framework operates as an end-to-end solution requiring no specialized expertise for clinical deployment once trained. In scenarios where imaging protocols remain fixed, the trained model can be directly implemented without modification. Protocol modification necessitates model retraining with updated synthetic data, but this process can be completed efficiently without clinical intervention. This combination of flexibility and operational simplicity positions SynQ-Breast well for clinical integration.

There are several limitations in this study. First, this study is based on single-center design and moderate sample size. Although this is sufficient for initial validation of the new method, multi-center studies with larger, more diverse cohorts are necessary in the future to validate comprehensive generalizability. Second, while this study focused on IVIM parameters to demonstrate the core capability of the proposed framework, we did not account for potential confounders, including patient demographic and clinical factors (e.g., menopausal status and hormone use) or other imaging-related factors (e.g., lesion size, margin morphology, and background parenchymal enhancement). These variables may influence the diagnostic performance of breast MRI. Third, the imaging characteristics of breast lesions (e.g., morphology, size, and enhancement characteristics) were not evaluated in this study, because the primary aim was to determine whether improved IVIM parameter estimation by SynQ-Breast could enhance benign/malignant classification. Fourth, while U-Net provides a robust and well-validated architecture for this initial implementation, emerging network designs offer promising avenues to improve the effectiveness of our framework. Transformer-based architectures and attention mechanisms may better capture long-range spatial dependencies in parameter maps, which is worth exploring. Fifth, the conventional IVIM bi-exponential model does not account for compartment-specific T₂ relaxation effects, which may contribute to relatively high estimates of f and low estimates of D*. Previous work showed that relaxation-corrected IVIM fitting reduced the estimated f compared with IVIM bi-exponential fitting and could improve the stability of D* estimation (39). Future studies incorporating relaxation-corrected IVIM model may help improve the reliability and physiological interpretability of IVIM parameters.

Conclusions

The SynQ-Breast framework successfully overcomes the critical challenges of data scarcity and noise sensitivity that have long hampered robust IVIM-MRI parameter estimation in breast tumors. By leveraging a novel synthetic data generation strategy, our method provides better parametric map quality, enhanced diagnostic performance in tumor differentiation, and a great reduction in computational time, in comparison to conventional NLLS method. This work establishes a powerful, dMRI-model-agnostic strategy for quantitative breast imaging, paving the way for SynQ-Breast to achieve more reliable and accessible clinical applications.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0275/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0275/dss

Funding: This work was supported by Guangdong Basic and Applied Basic Research Foundation (grant number 2024A1515011262), National Natural Science Foundation of China (grant number 12375291), National Key R&D Program of China (grant number 2023YFA1607502), and Research Initiation Fund of Jimei University (grant number ZQ2024085).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0275/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Medical Ethics Committee of The First Affiliated Hospital of Zhengzhou University (No. 2019-KY-231). Informed consent was obtained from all individual participants included in the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, Ferlay J, Siegel RL, Soerjomataram I, Jemal A. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Ginsburg O, Yip CH, Brooks A, Cabanes A, Caleffi M, Dunstan Yataco JA, et al. Breast cancer early detection: A phased approach to implementation. Cancer 2020;126:2379-93. [Crossref] [PubMed]
3. Xu M, Yan Y, Chen Y, Chen X, Gong K, Fu F. Expression and Subtype Discordance Between Core Needle Biopsy and Surgical Specimen in Breast Cancer. J Surg Res 2025;307:42-52. [Crossref] [PubMed]
4. Lo Gullo R, Sevilimedu V, Baltzer P, Le Bihan D, Camps-Herrero J, Clauser P, Gilbert FJ, Iima M, Mann RM, Partridge SC, Patterson A, Sigmund EE, Thakur S, Thibault FE, Martincich L, Pinker K. EUSOBI International Breast Diffusion-Weighted Imaging working group. A survey by the European Society of Breast Imaging on the implementation of breast diffusion-weighted imaging in clinical practice. Eur Radiol 2022;32:6588-97. [Crossref] [PubMed]
5. Sun K, Chen X, Chai W, Fei X, Fu C, Yan X, Zhan Y, Chen K, Shen K, Yan F. Breast Cancer: Diffusion Kurtosis MR Imaging-Diagnostic Accuracy and Correlation with Clinical-Pathologic Factors. Radiology 2015;277:46-55. [Crossref] [PubMed]
6. Zhang X, Teng X, Zhang J, Lai Q, Cai J. Enhancing pathological complete response prediction in breast cancer: the role of dynamic characterization of DCE-MRI and its association with tumor heterogeneity. Breast Cancer Res 2024;26:77. [Crossref] [PubMed]
7. Sun SY, Ding Y, Li Z, Nie L, Liao C, Liu Y, Zhang J, Zhang D. Multiparameter MRI Model With DCE-MRI, DWI, and Synthetic MRI Improves the Diagnostic Performance of BI-RADS 4 Lesions. Front Oncol 2021;11:699127. [Crossref] [PubMed]
8. Pinker K, Moy L, Sutton EJ, Mann RM, Weber M, Thakur SB, Jochelson MS, Bago-Horvath Z, Morris EA, Baltzer PA, Helbich TH. Diffusion-Weighted Imaging With Apparent Diffusion Coefficient Mapping for Breast Cancer Detection as a Stand-Alone Parameter: Comparison With Dynamic Contrast-Enhanced and Multiparametric Magnetic Resonance Imaging. Invest Radiol 2018;53:587-95. [Crossref] [PubMed]
9. Daimiel Naranjo I, Gibbs P, Reiner JS, Lo Gullo R, Sooknanan C, Thakur SB, Jochelson MS, Sevilimedu V, Morris EA, Baltzer PAT, Helbich TH, Pinker K. Radiomics and Machine Learning with Multiparametric Breast MRI for Improved Diagnostic Accuracy in Breast Cancer Diagnosis. Diagnostics (Basel) 2021;11:919. [Crossref] [PubMed]
10. Daimiel Naranjo I, Gibbs P, Reiner JS, Lo Gullo R, Thakur SB, Jochelson MS, Thakur N, Baltzer PAT, Helbich TH, Pinker K. Breast Lesion Classification with Multiparametric Breast MRI Using Radiomics and Machine Learning: A Comparison with Radiologists' Performance. Cancers (Basel) 2022;14:1743. [Crossref] [PubMed]
11. Jelescu IO, Veraart J, Fieremans E, Novikov DS. Degeneracy in model parameter estimation for multi-compartmental diffusion in neuronal tissue. NMR Biomed 2016;29:33-47. [Crossref] [PubMed]
12. Graham SJ, Stanchev PL, Lloyd-Smith JO, Bronskill MJ, Plewes DB. Changes in fibroglandular volume and water content of breast tissue during the menstrual cycle observed by MR imaging at 1.5 T. J Magn Reson Imaging 1995;5:695-701. [Crossref] [PubMed]
13. Kuhl CK, Kooijman H, Gieseke J, Schild HH. Effect of B1 inhomogeneity on breast MR imaging at 3.0 T. Radiology 2007;244:929-30. [Crossref] [PubMed]
14. Marino MA, Helbich T, Baltzer P, Pinker-Domenig K. Multiparametric MRI of the breast: A review. J Magn Reson Imaging 2018;47:301-15. [Crossref] [PubMed]
15. McDougall MP, Cheshkov S, Rispoli J, Malloy C, Dimitrov I, Wright SM. Quadrature transmit coil for breast imaging at 7 tesla using forced current excitation for improved homogeneity. J Magn Reson Imaging 2014;40:1165-73. [Crossref] [PubMed]
16. Azlan CA, Di Giovanni P, Ahearn TS, Semple SI, Gilbert FJ, Redpath TW. B1 transmission-field inhomogeneity and enhancement ratio errors in dynamic contrast-enhanced MRI (DCE-MRI) of the breast at 3T. J Magn Reson Imaging 2010;31:234-9. [Crossref] [PubMed]
17. Sung K, Daniel BL, Hargreaves BA. Transmit B1+ field inhomogeneity and T1 estimation errors in breast DCE-MRI at 3 tesla. J Magn Reson Imaging 2013;38:454-9. [Crossref] [PubMed]
18. Golkov V, Dosovitskiy A, Sperl JI, Menzel MI, Czisch M, Samann P, Brox T, Cremers D. q-Space Deep Learning: Twelve-Fold Shorter and Model-Free Diffusion MRI Scans. IEEE Trans Med Imaging 2016;35:1344-51. [Crossref] [PubMed]
19. Ye C, Cui Y, Li X. q-space learning with synthesized training data. In: Bonet-Carne E, Grussu F, Ning L, Sepehrband F, Tax CMW. editors. Computational Diffusion MRI. Cham: Springer International Publishing, 2019:123-32.
20. Barbieri S, Gurney-Champion OJ, Klaassen R, Thoeny HC. Deep learning how to fit an intravoxel incoherent motion model to diffusion-weighted MRI. Magn Reson Med 2020;83:312-21. [Crossref] [PubMed]
21. Valindria V, Palombo M, Chiou E, Singh S, Punwani S, Panagiotaki E. Synthetic q-space learning with deep regression networks for prostate cancer characterisation with verdict. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). 2021:50-4.
22. Kaandorp MPT, Barbieri S, Klaassen R, van Laarhoven HWM, Crezee H, While PT, Nederveen AJ, Gurney-Champion OJ. Improved unsupervised physics-informed deep learning for intravoxel incoherent motion modeling and evaluation in pancreatic cancer patients. Magn Reson Med 2021;86:2250-65. [Crossref] [PubMed]
23. Vasylechko SD, Warfield SK, Afacan O, Kurugol S. Self-supervised IVIM DWI parameter estimation with a physics based forward model. Magn Reson Med 2022;87:904-14. [Crossref] [PubMed]
24. Kaandorp MPT, Zijlstra F, Karimi D, Gholipour A, While PT. Neighborhood-attention models for incorporating spatial information in deep learning parameter estimation applied to IVIM. In: Proceedings of the International Society for Magnetic Resonance in Medicine. Singapore; 2024:1139.
25. Masutani Y, Konya K, Kato E, Mori N, Ota H, Mugikura S, Takase K, Ichinoseki Y. Synthetic X-Q space learning for diffusion MRI parameter estimation: a pilot study in breast DKI. Int J Comput Assist Radiol Surg 2025;20:2423-35. [Crossref] [PubMed]
26. Kaandorp MPT, Zijlstra F, Karimi D, Gholipour A, While PT. Incorporating spatial information in deep learning parameter estimation with application to the intravoxel incoherent motion model in diffusion-weighted MRI. Med Image Anal 2025;101:103414. [Crossref] [PubMed]
27. Cohen O, Zhu B, Rosen MS. MR fingerprinting Deep RecOnstruction NEtwork (DRONE). Magn Reson Med 2018;80:885-94. [Crossref] [PubMed]
28. Kaandorp MPT, Zijlstra F, Federau C, While PT. Deep learning intravoxel incoherent motion modeling: Exploring the impact of training features and learning strategies. Magn Reson Med 2023;90:312-28. [Crossref] [PubMed]
29. Bertleff M, Domsch S, Weingärtner S, Zapp J, O'Brien K, Barth M, Schad LR. Diffusion parameter mapping with the combined intravoxel incoherent motion and kurtosis model using artificial neural networks at 3 T. NMR Biomed 2017;30:e3833. [Crossref] [PubMed]
30. Chen L, Schär M, Chan KWY, Huang J, Wei Z, Lu H, Qin Q, Weiss RG, van Zijl PCM, Xu J. In vivo imaging of phosphocreatine with artificial neural networks. Nat Commun 2020;11:1072. [Crossref] [PubMed]
31. Liu F, Kijowski R, El Fakhri G, Feng L. Magnetic resonance parameter mapping using model-guided self-supervised deep learning. Magn Reson Med 2021;85:3211-26. [Crossref] [PubMed]
32. Ottens T, Barbieri S, Orton MR, Klaassen R, van Laarhoven HWM, Crezee H, Nederveen AJ, Zhen X, Gurney-Champion OJ. Deep learning DCE-MRI parameter estimation: Application in pancreatic cancer. Med Image Anal 2022;80:102512. [Crossref] [PubMed]
33. Sigmund EE, Cho GY, Kim S, Finn M, Moccaldi M, Jensen JH, Sodickson DK, Goldberg JD, Formenti S, Moy L. Intravoxel incoherent motion imaging of tumor microenvironment in locally advanced breast cancer. Magn Reson Med 2011;65:1437-47. [Crossref] [PubMed]
34. Suo S, Cheng F, Cao M, Kang J, Wang M, Hua J, Hua X, Li L, Lu Q, Liu J, Xu J. Multiparametric diffusion-weighted imaging in breast lesions: Association with pathologic diagnosis and prognostic factors. J Magn Reson Imaging 2017;46:740-50. [Crossref] [PubMed]
35. Lee YJ, Kim SH, Kang BJ, Kang YJ, Yoo H, Yoo J, Lee J, Son YH, Grimm R. Intravoxel incoherent motion (IVIM)-derived parameters in diffusion-weighted MRI: Associations with prognostic factors in invasive ductal carcinoma. J Magn Reson Imaging 2017;45:1394-406. [Crossref] [PubMed]
36. Cho GY, Moy L, Kim SG, Baete SH, Moccaldi M, Babb JS, Sodickson DK, Sigmund EE. Evaluation of breast cancer using intravoxel incoherent motion (IVIM) histogram analysis: comparison with malignant status, histological subtype, and molecular prognostic factors. Eur Radiol 2016;26:2547-58. [Crossref] [PubMed]
37. Xu FY, Hu GW, Zheng CJ, Huang H, Luo JQ, Zheng GP, Duan XH, Loffroy R, Wáng YXJ. Increased Brownian Motion of Water Molecules in Livers With Advanced Fibrosis: Preclinical and Clinical Observations With Magnetic Resonance Imaging (MRI). Mol Imaging 2026;25:15353508261465601. [Crossref] [PubMed]
38. Li XM, Yao DQ, Quan XY, Li M, Chen W, Wáng YXJ. Perfusion of hepatocellular carcinomas measured by diffusion-derived vessel density biomarker: Higher hepatocellular carcinoma perfusion than earlier intravoxel incoherent motion reports. NMR Biomed 2024;37:e5125. [Crossref] [PubMed]
39. Jerome NP, d'Arcy JA, Feiweier T, Koh DM, Leach MO, Collins DJ, Orton MR. Extended T2-IVIM model for correction of TE dependence of pseudo-diffusion volume fraction in clinical diffusion-weighted magnetic resonance imaging. Phys Med Biol 2016;61:N667-80.
40. Li Z, Fan Q, Bilgic B, Wang G, Wu W, Polimeni JR, Miller KL, Huang SY, Tian Q. Diffusion MRI data analysis assisted by deep learning synthesized anatomical images (DeepAnat). Med Image Anal 2023;86:102744. [Crossref] [PubMed]
41. Freiman M, Perez-Rossello JM, Callahan MJ, Voss SD, Ecklund K, Mulkern RV, Warfield SK. Reliable estimation of incoherent motion parametric maps from diffusion-weighted MRI using fusion bootstrap moves. Med Image Anal 2013;17:325-36. [Crossref] [PubMed]
42. Vidić I, Jerome NP, Bathen TF, Goa PE, While PT. Accuracy of breast cancer lesion classification using intravoxel incoherent motion diffusion-weighted imaging is improved by the inclusion of global or local prior knowledge with bayesian methods. J Magn Reson Imaging 2019;50:1478-88. [Crossref] [PubMed]
43. Jin YN, Zhang Y, Cheng JL, Zhang XP, Hu Y, Shao XN. The role of histogram analysis in diffusion-weighted imaging in the differential diagnosis of benign and malignant breast lesions. BMC Med Inform Decis Mak 2020;20:239. [Crossref] [PubMed]
44. Kaandorp MPT, Jakab A, Federau C, While PT. A Comparative Study of IVIM-MRI Fitting Techniques in Glioma Grading: Conventional, Bayesian, and Voxel-Wise and Spatially-Aware Deep Learning Approaches. J Magn Reson Imaging 2026;64:184-200. [Crossref] [PubMed]
45. Huang H, Liu B, Xu Y, Zhou W. Synthetic-to-real domain adaptation with deep learning for fitting the intravoxel incoherent motion model of diffusion-weighted imaging. Med Phys 2023;50:1614-22. [Crossref] [PubMed]
46. Yuan J, Wong OL, Lo GG, Chan HH, Wong TT, Cheung PS. Statistical assessment of bi-exponential diffusion weighted imaging signal characteristics induced by intravoxel incoherent motion in malignant breast tumors. Quant Imaging Med Surg 2016;6:418-29. [Crossref] [PubMed]
47. Cho GY, Moy L, Zhang JL, Baete S, Lattanzi R, Moccaldi M, Babb JS, Kim S, Sodickson DK, Sigmund EE. Comparison of fitting methods and b-value sampling strategies for intravoxel incoherent motion in breast cancer. Magn Reson Med 2015;74:1077-85. [Crossref] [PubMed]
48. Dixon WT. Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging: a modest proposal with tremendous potential. Radiology 1988;168:566-7. [Crossref] [PubMed]