Abstract
Background
To develop a deep learning segmentation method utilizing the nnU-Net architecture for fully automated whole-breast segmentation based on diffusion-weighted imaging (DWI) and synthetic MRI (SyMRI) images.
Methods
A total of 98 patients with 196 breasts were evaluated. All patients underwent 3.0T magnetic resonance (MR) examinations, which incorporated DWI and SyMRI techniques. The ground truth for breast segmentation was established through a manual, slice-by-slice approach performed by two experienced radiologists. The U-Net and nnU-Net deep learning algorithms were employed to segment the whole-breast. Performance was evaluated using various metrics, including the Dice Similarity Coefficient (DSC), accuracy, and Pearson’s correlation coefficient.
Results
For DWI and proton density (PD) of SyMRI, the nnU-Net outperformed the U-Net achieving the higher DSC in both the testing set (DWI, 0.930 ± 0.029 vs. 0.785 ± 0.161; PD, 0.969 ± 0.010 vs. 0.936 ± 0.018) and independent testing set (DWI, 0.953 ± 0.019 vs. 0.789 ± 0.148; PD, 0.976 ± 0.008 vs. 0.939 ± 0.018). The PD of SyMRI exhibited better performance than DWI, attaining the highest DSC and accuracy. The correlation coefficients R² for nnU-Net were 0.99 ~ 1.00 for DWI and PD, significantly surpassing the performance of U-Net.
Conclusion
The nnU-Net exhibited exceptional segmentation performance for fully automated breast segmentation of contrast-free quantitative images. This method serves as an effective tool for processing large-scale clinical datasets and represents a significant advancement toward computer-aided quantitative analysis of breast DWI and SyMRI images.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12880-025-01928-2.
Keywords: Synthetic magnetic resonance imaging, Diffusion-weighted imaging, Whole-breast segmentation, nnU-Net, U-Net
Introduction
Breast cancer remains the most frequently diagnosed cancer and the second leading cause of death in the female population worldwide [1]. Early detection coupled with precise diagnosis and appropriate treatment selection is crucial to enhancing patient survival rates. Magnetic resonance imaging (MRI) has become an indispensable tool in clinical practice for high-risk screening, diagnosis, preoperative staging, and evaluation of response to neoadjuvant therapy [2, 3].
Dynamic contrast-enhanced (DCE) MRI, traditionally considered the initial MRI technique for breast cancer evaluation, requires intravenous administration of contrast agents and relatively long acquisition times. Recently, contrast-free quantitative MRI methods have gained increasing attention due to their non-invasive nature, reduced acquisition times, and promising diagnostic potential. Among these, diffusion-weighted imaging (DWI), which assesses water molecule diffusion characteristics, and synthetic MRI (SyMRI), which rapidly provides quantitative tissue parameters [T1 and T2 relaxation times, proton density (PD)], have emerged as valuable alternatives or complementary modalities to DCE-MRI [4–9]. These quantitative MRI techniques have demonstrated considerable potential in breast cancer diagnosis [6, 7], molecular subtyping, prognostic prediction [8, 9], and assessing therapeutic efficacy [10].
A critical prerequisite step for quantitative breast MRI analysis is the accurate and robust segmentation of the whole breast region. Whole-breast segmentation enables the computation of quantitative imaging biomarkers, multimodal image registration, tumor segmentation, background parenchymal enhancement evaluation, and breast density measurement—each closely associated with breast cancer risk stratification and prognosis [11–14]. However, accurate whole-breast segmentation remains a challenging and labor-intensive task due to the anatomical complexity, variability in breast shape and size, and large imaging datasets generated during MRI examinations.
In recent years, some studies based on deep learning algorithms have been conducted to automatically segment the whole-breast [15–17]. Among these methods, the widely used U-Net convolutional neural network (CNN) architecture and its variants have demonstrated superior segmentation performance across various breast MRI modalities [18, 19]. However, some extra pre-and/or post-processing algorithms, which are complex and time-consuming, may be required. Recently, nnU-Net—a robust, self-configuring deep learning segmentation framework—has demonstrated superior segmentation performance compared to conventional U-Net-based architectures in various medical imaging segmentation tasks [20, 21]. Nevertheless, to the best of our knowledge, few researchers have investigated fully automated methods for such segmentation on SyMRI of the breast, and much less have we seen literature comparing the segmentation results of the two quantitative, contrast-free methods of DWI and SyMRI.
The purpose of this study was to develop a deep learning segmentation method utilizing the nnU-Net architecture, initially for breast segmentation within whole images, followed by an assessment of the performance of fully automated segmentation based on DWI and SyMRI images.
Materials and methods
Datasets
This retrospective study received approval from our institutional review board, which granted a waiver for informed consent. Patients were excluded if they had undergone any type of surgery, such as partial or total mastectomy, implant placement, or if they presented with large and protruding lesions. A cohort of 98 consecutive patients with both benign and malignant breast lesions (comprising 196 breasts; median age: 48 years; age range: 24–74 years) was enrolled at our institution as illustrated in Fig. 1. All participants underwent SyMRI and DWI of the breast from May 2019 to October 2019. MRI was conducted using a 3-T whole-body scanner (Signa Pioneer, GE Healthcare, Milwaukee, WI). The dataset included DWI sequences and PD mappings derived from SyMRI, which were utilized to develop and assess our segmentation methodologies. Explicit scan parameters of SyMRI (MAGiC) and DWI are presented in Table Appendix 1.
Fig. 1.
The flowchart of patient enrollment
Ground truth segmentation
The segmentation task was conducted independently by two radiologists (Y.Y.Z., B.G.) using ITK-SNAP version 3.8 [22] in a manual slice-by-slice approach, with both radiologists blinded to each other’s assessments to maintain an unbiased evaluation of the images. Subsequently, the segmentations were reviewed by an expert radiologist (W.B.G.). The results deemed accurate by the expert radiologist were established as the ground truth for neural network training and for the objective quantification of segmentation performance.
Whole-breast segmentation based on U-Net and nnU-Net
Given the variability in slice thickness of the DWI and SyMRI sequences, we employed a 2D configuration for both U-Net and nnU-Net architectures to train whole breast segmentation models in this study. As illustrated in Appendix 2, U-Net employs an encoder-decoder architecture featuring skip connections. The blue rectangles represent feature maps, with the number of channels specified at the top of each rectangle and the x-y dimensions indicated at the bottom-left corner. The encoder on the left was tasked with feature extraction, while the decoder on the right was responsible for reconstructing the encoded information. By leveraging the information obtained through the skip connections, U-Net was capable of directly incorporating spatial data into its predictions.
For the U-Net architecture, the following parameters were employed: a training batch size of 8, the Adam optimizer, and an initial learning rate of 0.001. Identifying optimal hyperparameters for U-Net models necessitates extensive experimentation by seasoned researchers. In this context, nnU-Net diverges from conventional methods not by modifying the U-Net architecture, but by automating the optimization process for training parameters. Initially, nnU-Net processes the dataset to generate dataset fingerprints, which encompass characteristics such as image size and modality. Subsequently, it configures parameters automatically, such as batch size and patch size, based on predefined rules. These parameters are then integrated with established blueprint parameters, including learning rate and loss functions, to produce comprehensive pipeline fingerprints.
The nnU-Net [23], after analyzing various breast DWI and SyMRI sequence images, determined all necessary parameters for training both whole-breast segmentation models in 2D patterns and subsequently adopted these parameters for the neural network training. The nnU-Net incorporated parameter search assignments. Thus, it is not required to define loss functions and optimizers. Subsequently, to train the nnU-Net, the source code offered by the authors was employed. The study employed a rigorous evaluation partitioning strategy with 98 patients partitioned into three subsets: 71 for training set and 17 for testing set (5-fold cross-validation), 10 as an independent testing set. 5-Fold Cross-Validation ensured sufficient learning capacity while maintaining statistical validity. The independent testing set evaluated model generalizability to unseen clinical environments. For each model, a 5-fold cross-validation method with the same seed number. The 5-fold cross-validation was subsequently adopted to streamline the hyperparameters. Besides, the optimal model was employed in an independent testing set of 10 patients. All models were trained using Python 3.7 and PyTorch 1.7.1 (Appendix 3).
Statistical analysis
We employed five widely used metrics to evaluate the performance of whole-breast segmentation: the dice similarity coefficient (DSC), accuracy, sensitivity, specificity, and average symmetric surface distance (ASD). DSC quantifies the similarity between predicted labels and ground truth, while accuracy assesses the overall effectiveness of the segmentation model. Sensitivity and specificity measure the under-segmentation ratio, and ASD calculates the average distance between two surfaces reconstructed from the predicted and ground-truth labels. Higher values of DSC, accuracy, sensitivity, and specificity, along with lower values of ASD, indicate superior segmentation performance. Metric computation adopted a two-tiered strategy: slice-level metrics were first calculated through pixel-wise comparisons on individual 2D slices, followed by patient-level aggregation using volume-weighted averaging across all slices to preserve spatial coherence. Segmentation outcomes for U-Net, nnU-Net, and manual methods were compared using mean ± standard deviation values of all metrics, with superior performance indicated by higher DSC/accuracy/sensitivity/specificity and lower ASD values. Moreover, an anatomical parameter such as volume was included as a clinically applicable metric. Pearson’s correlation coefficient (R², 0–1) was used to assess the correlation between the prediction outputs of U-Net and nnU-Net and the ground truth volume, with higher R² values reflecting stronger correlation. To ensure robustness, we also assessed the consistency between the manual annotations provided by two radiologists. This dual analysis is instrumental in evaluating both the reliability of the automated models and the variability inherent in manual annotations among radiologists.
Results
Interobserver agreement of whole-breast segmentation
The concordance between manual segmentations performed by the two radiologists exhibited a high degree of correlation. Specifically, the DSC for spatial consistency was 0.970 for PD images and 0.917 for DWI images. The metrics of accuracy, sensitivity, specificity, and ASD are presented in Table 1 and illustrated in Fig. 2.
Table 1.
Correlation of whole breast segmentation by two radiologists
| PD | DWI | |
|---|---|---|
| DSC | 0.970 ± 0.021 | 0.917 ± 0.065 |
| Accuracy | 0.995 ± 0.004 | 0.989 ± 0.008 |
| Sensitivity | 0.991 ± 0.012 | 0.932 ± 0.074 |
| Specificity | 0.995 ± 0.005 | 0.993 ± 0.005 |
| ASD (mm) ↓ | 6.966 ± 1.143 | 5.971 ± 1.038 |
Data are mean ± standard deviation. ASD, average symmetric surface distance; DSC, dice similarity coefficient; DWI, diffusion-weighted imaging; PD, proton density
Fig. 2.
Correlation of whole breast segmentation between the two radiologists. DWI (a), PD (b). The red line is the trend line, and the dashed black line is the unity line as reference
Breast segmentation on DWI images
Representative segmentation images derived from DWI are presented in Fig. 3. The values for the DSC, accuracy, sensitivity, specificity, and ASD are summarized in Table 2 (testing set) and Table 3 (independent testing set). The nnU-Net model demonstrated significantly higher DSC, accuracy, sensitivity, and specificity compared to the U-Net model in both the testing set and independent testing set, while exhibiting lower ASD values. Notably, the nnU-Net achieved the highest DSC (0.953 ± 0.019) and accuracy (0.993 ± 0.002) in the independent testing set. The correlation coefficient R² for nnU-Net was 0.99 in both the testing set and independent testing set, surpassing that of the U-Net, thereby indicating a strong agreement between predicted outputs and ground truth (Figs. 4 and 5). The trend line slope approached 1, further supporting this conclusion.
Fig. 3.
Segmentation results of DWI images from four different patients using nnU-Net and U-Net on independent testing set. The nnU-Net model outperformed the U-Net in this assessment. DWI, diffusion-weighted imaging. The arrows indicate that compared to the ground truth, U-net performs significantly worse than nnU-net in segmentation involving larger breast margins, near the chest wall, close to the thoracic cavity (ribs), and regions with artifacts
Table 2.
The segmentation performance for whole breast in testing set
| Method | Testing set (n = 17) | ||||
|---|---|---|---|---|---|
| DSC | Accuracy | Sensitivity | Specificity | ASD (mm) ↓ | |
| PD_U-Net | 0.936 ± 0.018 | 0.991 ± 0.004 | 0.914 ± 0.032 | 0.997 ± 0.002 | 9.167 ± 1.524 |
| PD_nnU-Net * | 0.969 ± 0.010 | 0.996 ± 0.002 | 0.969 ± 0.017 | 0.998 ± 0.001 | 7.623 ± 1.519 |
| DWI_U-Net | 0.785 ± 0.161 | 0.976 ± 0.010 | 0.768 ± 0.202 | 0.990 ± 0.004 | 7.464 ± 0.446 |
| DWI_nnU-Net * | 0.930 ± 0.029 | 0.991 ± 0.004 | 0.935 ± 0.029 | 0.995 ± 0.003 | 6.494 ± 0.552 |
Data are mean ± standard deviation. ASD, average symmetric surface distance; DSC, dice similarity coefficient; DWI, diffusion-weighted imaging; PD, proton density. *Data performs better
Table 3.
The segmentation performance for whole breast in independent testing set
| Method | Independent testing set (n = 10) | ||||
|---|---|---|---|---|---|
| DSC | Accuracy | Sensitivity | Specificity | ASD (mm) ↓ | |
| PD_U-Net | 0.939 ± 0.018 | 0.990 ± 0.003 | 0.910 ± 0.034 | 0.998 ± 0.001 | 9.157 ± 1.495 |
| PD_nnU-Net * | 0.976 ± 0.008 | 0.996 ± 0.002 | 0.970 ± 0.012 | 0.998 ± 0.001 | 6.919 ± 1.080 |
| DWI_U-Net | 0.789 ± 0.148 | 0.971 ± 0.019 | 0.778 ± 0.157 | 0.986 ± 0.021 | 7.290 ± 0.747 |
| DWI_nnU-Net * | 0.953 ± 0.019 | 0.993 ± 0.002 | 0.952 ± 0.028 | 0.996 ± 0.001 | 6.114 ± 0.738 |
Data are mean ± standard deviation. ASD, average symmetric surface distance; DSC, dice similarity coefficient; DWI, diffusion-weighted imaging; PD, proton density. *Data performs better
Fig. 4.
Correlation of the segmentation between the ground truth obtained by and the U-Net prediction. Testing set DWI (a) and PD (c). Independent testing set DWI (b) and PD (d). The red line is the trend line, and the dashed black line is the unit line as a reference. DWI, diffusion-weighted imaging. PD, proton density
Fig. 5.
Correlation of the segmentation between the ground truth obtained by and the nnU-Net prediction. Testing set DWI (a) and PD (c). Independent testing set DWI (b) and PD (d). The red line is the trend line, and the dashed black line is the unit line as a reference. DWI, diffusion-weighted imaging. PD, proton density
Breast segmentation on SyMRI (PD) images
As illustrated in Appendix 4, after comparing the three quantitative maps generated by SyMRI (T1, T2, and PD maps), the breast boundaries in the T1 and T2 maps were found to be less distinct than those in the PD map, making accurate manual segmentation of the entire breast region impractical on T1 and T2 maps. Therefore, the PD maps were selected as the optimal imaging modality for performing whole-breast segmentation in this study. Representative segmentation images derived from PD imaging are illustrated in Fig. 6. The metrics for DSC, accuracy, sensitivity, specificity, and ASD are summarized in Table 2 (testing set) and Table 3 (independent testing set). The nnU-Net demonstrated markedly higher values for DSC, accuracy, sensitivity, and specificity compared to the U-Net model across both the testing set and independent testing set, while the ASD values for nnU-Net were lower than those for U-Net. In the independent testing set, nnU-Net achieved the highest DSC of 0.976 ± 0.008 and an accuracy of 0.996 ± 0.002. Furthermore, the correlation coefficients R² for nnU-Net were 0.99 and 1.00 for the testing set and independent testing set, respectively, which significantly surpassed the performance of U-Net. This indicates a strong concordance between the predicted outputs and the ground truth (Figs. 4 and 5).
Fig. 6.
Segmentation results of PD images from four different patients (identical to the DWI images in Fig. 3) using nnU-Net and U-Net on independent test set. The nnU-Net model outperformed the U-Net in this assessment. DWI, diffusion-weighted imaging. PD, proton density. The arrow indicates that, compared to the ground truth, U-net performs slightly worse than nnU-net in segmenting larger breasts and those closer to the chest wall
Time of the segmentation
The average time for manual segmentation of DWI and PD images by two radiologists was more than 25 min and 20 min per MRI volume, respectively. In the testing set, the average time for U-Net and nnU-Net took about 6.48 s and 4.44 s, respectively, to perform the whole-breast segmentation for DWI images, and about 6.33 s and 3.90 s, respectively, to perform the whole segmentation for PD images. In the independent testing set, the average time for U-Net and nnU-Net took about 6.33 s and 4.37 s, respectively, to perform the whole-breast segmentation for DWI images, and about 6.08 s and 4.05 s, respectively, to perform the whole segmentation for PD images.
Discussion
The successful implementation of fully whole-breast segmentation, which would facilitate the subsequent automated detection and segmentation of tumors, could enable accurate assessment of benign or malignant lesions through quantitative parameters. This advancement holds the promise of achieving a fully automated and enhancement-free breast MRI screening that is efficient and rapid. Such a methodology would serve as a significant complement or potential alternative to existing breast cancer screening modalities reliant on mammography and ultrasound technologies. However, whole-breast segmentation is a highly time-consuming and labor-intensive task, making the quantitative calculation of these metrics in clinical practice extremely challenging due to the lack of suitable automated segmentation tools. Our current study investigated the application of U-Net and nnU-Net for whole-breast segmentation in quantitative DWI and SyMRI (PD), an area that has been rarely explored in previous research. The results demonstrated that SyMRI (PD) was superior to DWI images and that the nnU-Net outperformed the U-Net, achieving the highest DSC and accuracy. Consequently, our findings suggested that deep learning segmentation through nnU-Net was viable for completely automated segmentation of contrast-free quantitative MRI of breast with considerable accuracy in comparison with the ground truth of manual segmentation by radiologists.
Accurate and robust segmentation of the Whole-breast constitutes a fundamental prerequisite for quantitative analysis in breast MRI. Comprehensive whole-breast segmentation is essential for deriving quantitative imaging biomarkers, facilitating multimodal image registration, and enabling precise tumor segmentation. However, its integration into routine clinical workflows remains challenging due to the time-intensive nature of manual delineation. Our proposed method demonstrates a substantial reduction in processing time, operating approximately 200 times faster than manual segmentation performed by experienced radiologists. We believe this technique offers a clinically viable solution, particularly relevant in contemporary practice, where breast MRI is assuming an increasingly pivotal role in screening and diagnostic pathways.
Our results demonstrated that two fully automated deep learning methods, U-Net and nnU-Net, could be used in the whole-breast segmentation, corresponding to the previous observations [11, 14–17, 24, 25]. Zhang et al. [11] showed that U-Net could be used for the whole-breast segmentation of DWI images, which is consistent with our findings, but we increased the number of patients. Meanwhile, we found that another automatic deep learning algorithm, nnU-Net, significantly outperformed the U-Net in DWI images, achieving the DSC values of 0.930 ± 0.029 ∼ 0.953 ± 0.019 versus 0.690 ± 0.120 ∼ 0.85 ± 0.07. In addition, our study compared another quantitative MRI, SyMRI, which is currently in the spotlight. We found that U-Net and nnU-Net based on the PD images of SyMRI had better efficacy than DWI for whole-breast segmentation, with the DSC values of 0.936 ± 0.936 ∼ 0.939 ± 0.018 and 0.969 ± 0. 010 ∼ 0.976 ± 0.008, respectively, as well as very high sensitivity and specificity, and a low ASD value. Notably, our deep learning approach effectively mitigates the occurrence of over-segmentation while maintaining robust segmentation sensitivity.
Our findings strongly revealed the great potential of performing nnU-Net for breast segmentation, which aligns with the discoveries by Huo et al. [15]. Still, we obtained higher DSC values based on free contrast-enhanced images of SyMRI (PD) than the former based on DCE-MRI images (0.976 ± 0.008 versus 0.968 ± 0.017) with almost the same accuracy (0.996 ± 0.002 versus 0.996 ± 0.003). Our findings were also superior to the DSC values based on the free contrast-enhanced images of T1WI with or without fat suppression by Zhang et al. [15, 16] laying the groundwork for the widespread use of SyMRI (PD) in clinics. Above all, unlike former work in the literature on breast segmentation [11, 14–17, 24, 25], our performance metric values based on free contrast-enhanced PD images of SyMRI are impressive with higher DSC values, which may be more promising than traditional quantitative DWI images and non-quantitative T1WI images. However, larger multi-center studies are needed for validation.
Our analyses revealed that PD images of SyMRI outperform DWI images in the testing set and independent testing set, regardless of U-Net or nnU-Net automated whole-breast segmentation. Meanwhile, we also found that DWI images are even more challenging than PD images for whole-breast segmentation, due to the following reasons [11]: firstly, the DWI images are prone to artifacts, especially the outer contour edges of the breast, as shown in Fig. 3, look blurred, which greatly increases the difficulty of segmentation; secondly, the resolution between different tissues of the image is lower, especially between the breast and the pectoral muscle, and it is more difficult to identify the edges of the breasts close to the chest wall; and thirdly, the DWI image is prone to distortions, especially in the case of larger breasts. However, SyMRI with the multi-echo and multidelay acquisition method (MAGiC) can synchronously quantify tissues’ inherent magnetic attributes, including T1, T2, and PD maps, thus providing reliable information without suffering from the shortcomings of the abovementioned DWI images [6]. Among these maps, PD maps provide better contrast for certain tissues, helping to accurately identify and differentiate between different tissue types, such as fat, muscle, bone, etc., which can aid in the diagnosis and assessment of disease. Therefore, after pre-testing, we finally chose to segment the whole breast on PD images.
Throughout various imaging patterns, U-Net and nnU-Net have exhibited robust autosegmentation performance and anatomic structures. They are both the most common segmentation algorithms for biomedical imaging. nnU-Net extends the capabilities of U-Net by introducing a more robust framework for hyperparameter optimization and model configuration. This self-adaptive approach allows nnU-Net to automatically tailor its architecture and training process to different datasets without manual intervention. Additionally, nnU-Net incorporates advanced techniques such as dynamic patch cropping and improved data augmentation strategies, leading to enhanced segmentation accuracy and consistency across diverse medical imaging tasks.
In this study, nnU-Net demonstrated a consistently superior performance compared to U-Net in the context of breast segmentation. Several advantages of nnU-Net over U-Net were identified: firstly, the results achieved by nnU-Net closely approximated the ground truth for both DWI and PD images. Secondly, nnU-Net effectively identified various imaging challenges, including deformations in larger breasts observed in lower resolution DWI images, as well as in tissues with lower contrast, such as the breast and pectoral muscles. Notably, nnU-Net was also successful in delineating the boundaries between the breast and the chest wall, which are typically problematic and time-consuming. Thirdly, a detailed slice-by-slice comparison of the ground truth and nnU-Net results against the original images revealed instances where nnU-Net outperformed the established ground truth. Consistent with the findings reported by Huo et al. [15], the nnU-Net based approach mitigated the need for labor-intensive and complex parameter tuning through the implementation of a robust training strategy alongside adaptive preprocessing and postprocessing techniques. The framework presented in this study is designed for immediate application and can be seamlessly integrated into clinical workflows.
We observed that the disparities between the two deep learning segmentation methodologies and the ground truth established by manual segmentation primarily involve the delineation of the breast’s edges, particularly in the lateral region where the breast connects to the chest wall. This complex task is significantly constrained by the lack of consensus regarding the anatomical boundaries of the breast [16, 26]. Consequently, although the outcomes of deep learning segmentation may not align perfectly with those of manual segmentation, this discrepancy does not necessarily indicate flaws within the deep learning approach. The fully automated deep learning method offers an objective alternative that is not influenced by operator bias [14], serving as an efficient segmentation tool for large datasets and facilitating clinical applications.
Our research had some limitations. First, this was a single-center research. Future research should include multiple centers and a broader range of cases, such as postoperative patients. Second, we have only analyzed quantitative SyMRI (PD) and DWI images, and the developed segmentation algorithms may not directly apply to other non-quantitative images such as T1WI or DCE-MRI images. Third, the current U-Net and nnU-Net models are used for 2D segmentation, and we will explore using 3D convolution in the future. Lastly, the visual evaluation by radiologists was utilized, known as objective as the ground truth for segmentation.
In conclusion, we have demonstrated that both U-Net and nnU-Net are applicable to quantitative DWI and SyMRI (PD) images. Notably, nnU-Net, as an out-of-the-box deep learning approach, exhibits exceptional segmentation performance for fully automated breast segmentation of contrast-free quantitative images. This method serves as an effective tool for processing large-scale clinical datasets and represents a significant advancement toward computer-aided quantitative analysis of breast DWI and SyMRI images.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
Not applicable.
Abbreviations
- ASD
Average symmetric surface distance
- DSC
Dice similarity coefficient
- DWI
Diffusion-weighted imaging
- SyMRI
Synthetic magnetic resonance imaging
- PD
Proton density
Author contributions
Guarantors of integrity of entire study, W.B.G., Y.L.Z., Q.X.Y., Y.W.X., F.S.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, W.B.G., Y.Y.Z., Y.W.X.; clinical studies, Y.Y.Z., B.G., W.B.L., X.H.L.; statistical analysis, W.B.G., Y.Y.Z., T.H., X.S.; and manuscript editing, W.B.G., Y.L.Z., Q.X.Y., Y.W.X., F.S., G.X.H.
Funding
This study was supported by funding from Shaanxi Province Natural Science Basic Research Programme Project, Contract grant number: 2024JC-YBQN-0893; IIT Clinical Research Fund of The Second Affiliated Hospital of Xi’an Jiaotong University, Contract grant number: M019.
Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
This study was performed in accordance with the principles of the Declaration of Helsinki. The study was approved by the institutional review board of the Second Affiliated Hospital of Xi’an Jiaotong University, which granted a waiver for informed consent.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Weibo Gao and Yanyan Zhang contributed equally to this work and are co-first authors for this study.
References
- 1.Siegel RL, Giaquinto AN, Jemal A. Cancer statistics, 2024. CA Cancer J Clin. 2024;74(1):12–49. [DOI] [PubMed] [Google Scholar]
- 2.Mann RM, Athanasiou A, Baltzer PAT, et al. Breast cancer screening in women with extremely dense breasts recommendations of the European society of breast imaging (EUSOBI). Eur Radiol. 2022. [DOI] [PMC free article] [PubMed]
- 3.Mann RM, Cho N, Moy L. Breast MRI: state of the Art. Radiology. 2019;292(3):520–36. [DOI] [PubMed] [Google Scholar]
- 4.Zhang M, Horvat JV, Bernard DB, et al. Multiparametric MRI model with dynamic contrast-enhanced and diffusion-weighted imaging enables breast cancer diagnosis with high accuracy. J Magn Reson Imaging. 2019;49(3):864–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Sauer ST, Christner SA, Schlaiß T, et al. Diffusion-weighted breast MRI at 3 tesla: improved lesion visibility and image quality with a combination of water excitation and spectral fat saturation. Acad Radiol. 2023;30(9):1773–83. [DOI] [PubMed] [Google Scholar]
- 6.Gao W, Zhang S, Guo J, et al. Investigation of synthetic relaxometry and diffusion measures in the differentiation of benign and malignant breast lesions as compared to BI-RADS. J Magn Reson Imaging. 2021;53(4):1118–27. [DOI] [PubMed] [Google Scholar]
- 7.Gao W, Yang Q, Li X, et al. Quantitative assessment of breast tumor: comparison of four methods of positioning region of interest for synthetic relaxometry and diffusion measurement. Acad radiol. 2024;S1076-6332(24)00135-1. Online ahead of print. [DOI] [PubMed]
- 8.Gao W, Yang Q, Li X, et al. Synthetic MRI with quantitative mappings for identifying receptor status, proliferation rate, and molecular subtypes of breast cancer. Eur J Radiol. 2022;148:110168. [DOI] [PubMed] [Google Scholar]
- 9.Horvat J, Bernard-Davila B, Helbich T, et al. Diffusion-weighted imaging (DWI) with apparent diffusion coefficient (ADC) mapping as a quantitative imaging biomarker for prediction of immunohistochemical receptor status, proliferation rate, and molecular subtypes of breast cancer. J Magn Reson Imaging. 2019;50(3):836–46. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Du S, Gao S, Zhao R, et al. Contrast-free MRI quantitative parameters for early prediction of pathological response to neoadjuvant chemotherapy in breast cancer. Eur Radiol. 2022;32(8):5759–72. [DOI] [PubMed] [Google Scholar]
- 11.Zhang L, Mohamed AA, Chai R, et al. Automated deep learning method for whole-breast segmentation in diffusion-weighted breast MRI. J Magn Reson Imaging. 2020;51(2):635–43. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Zhang J, Cui Z, Zhou L, et al. Breast fibroglandular tissue segmentation for automated BPE quantification with iterative cycle-consistent semi-supervised learning. IEEE Trans Med Imaging. 2023;42(12):3944–55. [DOI] [PubMed] [Google Scholar]
- 13.Pérez-Benito FJ, Signol F, Perez-Cortes JC, et al. A deep learning system to obtain the optimal parameters for a threshold-based breast and dense tissue segmentation. Comput Methods Programs Biomed. 2020;195:105668. [DOI] [PubMed] [Google Scholar]
- 14.Zhang Y, Chen JH, Chang KT, et al. Automatic breast and fibroglandular tissue segmentation in breast MRI using deep learning by a fully-convolutional residual neural network U-Net. Acad Radiol. 2019;26(11):1526–35. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Huo L, Hu X, Xiao Q, et al. Segmentation of the whole breast and fibroglandular tissue using nnU-Net in dynamic contrast enhanced MR images. Magn Reson Imaging. 2021;82:31–41. [DOI] [PubMed] [Google Scholar]
- 16.Zhang Y, Chan S, Chen JH, et al. Development of U-Net breast density segmentation method for fat-sat MR images using transfer learning based on non-fat-sat model. J Digit Imaging. 2021;34(4):877–87. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Fashandi H, Kuling G, Lu Y, et al. An investigation of the effect of fat suppression and dimensionality on the accuracy of breast MRI segmentation using U-nets. Med Phys. 2019;46(3):1230–44. [DOI] [PubMed] [Google Scholar]
- 18.Soffer S, Ben-Cohen A, Shimon O, et al. Convolutional neural networks for radiologic images: a radiologist’s guide. Radiology. 2019;290(3):590–606. [DOI] [PubMed] [Google Scholar]
- 19.Dalmis¸ MU, Litjens G, et al. Using deep learning to segment breast and fibroglandular tissue in MRI volumes. Med Phys. 2017;44(2):533–46. [DOI] [PubMed] [Google Scholar]
- 20.Isensee F, Jaeger PF, Kohl SAA, et al. Automated design of deep learning methods for biomedical image segmentation. ArXiv. 2019;1–55. 10.1038/s41592-020-01008-z.
- 21.Heller N, Isensee F, Maier-Hein KH, et al. The state of the Art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: results of the KiTS19 challenge. Med Image Anal. 2021;67:101821. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Yushkevich PA, Pashchinskiy A, Oguz I, et al. User-guided segmentation of multi-modality medical imaging datasets with ITK-SNAP. Neuroinformatics. 2019;17(1):83–102. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Isensee F, Jaeger PF, Kohl SAA, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods. 2021;18(2):203–11. [DOI] [PubMed] [Google Scholar]
- 24.Piantadosi G, Sansone M, Fusco R, et al. Multi-planar 3D breast segmentation in MRI via deep convolutional neural networks. Artif Intell Med. 2020;103:101781. [DOI] [PubMed] [Google Scholar]
- 25.Ivanovska T, Jentschke TG, Daboul A, et al. A deep learning framework for efficient analysis of breast volume and fibroglandular tissue using MR data with strong artifacts. Int J Comput Assist Radiol Surg. 2019;14(10):1627–33. [DOI] [PubMed] [Google Scholar]
- 26.Nam Y, Park GE, Kang J, et al. Fully automatic assessment of background parenchymal enhancement on breast MRI using machine-learning models. J Magn Reson Imaging. 2021;53(3):818–26. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.