Fully Automated Deep Learning Method for Fibroglandular Tissue Segmentation in Breast MRI

doi:10.21203/rs.3.rs-1606703/v1

Download PDF

Article

Fully Automated Deep Learning Method for Fibroglandular Tissue Segmentation in Breast MRI

https://doi.org/10.21203/rs.3.rs-1606703/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Studies have demonstrated that the relative amount of fibroglandular tissue (FGT) with respect to fatty tissue on breast MRI is a risk factor for breast cancer. The amount of FGT is typically assessed qualitatively by a radiologist or by semi-quantitative tools, but both methods suffer from inter-reader variability and are imprecise. The additional challenge is a lack of publicly available data for development and comparison of quantitative tools. The objective of this study was: (i) to develop fully automated segmentation methods for breast and FGT based on convolutional neural networks (CNN) and (ii) to publicly share annotation masks and radiologists scores for the utilized images. 127 fat-saturated gradient echo T1-weighted pre-contrast MRI studies were used for model development and evaluation. The models were evaluated using the Dice similarity coefficient (DSC). Additionally, we implemented and evaluated an algorithm for breast density assessment which utilized our segmentation models. In the test dataset, the DSC was 0.879 for the breast and 0.730 for the FGT. The correlation between median radiologist breast density assessment and FGT percentage assessed by our model was r=0.83 (p<0.01). Overall, the model demonstrated high accuracy compared to gold standard radiologist assessment. The data and the model were made publicly available.

Breast cancer is the most commonly diagnosed cancer in the world, accounting for a quarter of all cancer cases among women.¹ Mammography is the standard of care in to detect breast cancer. However, mammography suffers from several limitations, importantly the reduced contrast between tumors and the surrounding tissue. Tissue superimposition reduces the sensitivity of mammography, especially in dense breasts.^2–4 Magnetic resonance imaging (MRI) is the most sensitive breast imaging modality and provides three-dimensional reconstructed visualization of tissues thus avoiding overlap.

Multiple studies have demonstrated that the relative amount of fibroglandular tissue (FGT) with respect to fatty tissue on breast MRI is a risk factor of breast cancer.^5–7 Although the amount of FGT is typically assessed qualitatively by a radiologist, quantitative tools can more precisely measure the FGT volume.^8–10 However, hundreds of imaging slices are acquired per MRI sequence which requires extensive segmentation for quantitative methods. The manual process of segmentation is subjective and time-intensive, making it impractical for large dataset. Alternatively, FGT can be assessed via user-assisted interactive techniques such as thresholding. Although these techniques are faster, they still are affected by inter-reader variability and are not precise enough to achieve a reliable quantification of the breast density volume. Hence, an efficient, objective, and reliable segmentation method is needed. Automated segmentation techniques have the potential to reduce subjectivity and speed up data processing. However, automated segmentation of the breast is a challenging task due to breast size/shape variations, image intensity inhomogeneity, image artifact, and other noise errors.

In recent years, deep learning algorithms have been widely applied to biomedical image classification and segmentation applications. The main advantage of the deep learning methods is their ability to learn and recognize relevant features directly from examples. Deep learning approaches do not require any changes in their feature designs or the use of filters specific for each problem. Particularly, U-Net is a convolutional neural network (CNN) that has been shown to be fast and precise in segmenting biomedical images. The purpose of this study is to develop and validate a fully automated deep-learning segmentation method based on the U-net architecture for breast and FGT segmentation and quantification.

2.1 Study population

The breast MRI images were obtained from the Duke Breast Cancer MRI dataset (publicly available online at https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70226903). The dataset is composed of a retrospectively collected cohort of 922 biopsy-confirmed invasive breast cancer patients from a single institution with preoperative MRI from January 1, 2000 to March 23, 2014. The MR images are available in DICOM format with non-fat saturated T1-weighted, fat-saturated gradient echo T1-weighted pre-contrast, and post-contrast sequences. An overview of the study population used in the training, validation, and test sets can be found in Table 1. Other relevant clinical data such as tumor characteristics, genomic data, and clinical outcomes are also included in the public dataset. More information on the dataset can be found in a previous publication.¹¹

Table 1

Patient Characteristics
Characteristic	Patients (n = 127)
Age – yrs	53.1 ± 10.8
Race – no (%)
• White	91 (71.6)
• Black	28 (22.0)
• Hispanic	3 (2.3)
• Other	5 (4.1)
Menopause Status – no (%)
• Positive	72 (56.7)
• Negative	53 (41.7)
• Not reported	2 (1.6)

2.2 Magnetic Resonance Imaging protocol

With the patients placed in a prone position, MRI was performed on 1.5- or 3.0T scanners (GE Healthcare and Siemens) to obtain axial breast images. Slice thickness was 1.04-2.5mm. Repetition time was 3.54–7.39 ms and echo time was 1.25–2.76 ms. The acquisition matrix had a minimum array size of 320 x 320 and maximum array size of 448 x 448. Flip angle was 7–12 degrees and field of view was 250–480 mm.

2.3 Image selection and annotation

From the Duke Breast Cancer MRI dataset, we selected 127 fat-saturated gradient echo T1-weighted pre-contrast MRI studies. For each study, three image slices were chosen for breast segmentation and three image slices were chosen for FGT segmentation. To guarantee the comprehensiveness of our MRI dataset, the following selection rules were applied: to select images for breast segmentation, the MRI volume was evenly divided into thirds from which one image slice was randomly selected from each third of the volume; to select images for FGT segmentation, the MRI study was evenly divided into fourths from which two image slices were randomly selected from the middle half (since most fibroglandular tissue is concentrated in the middle of the breast) and one image slice was randomly selected from the first or last quarter sections of the MRI study.

For each MRI image, the outer contours of the breast and the FGT were traced. Manual segmentation was performed on the selected images to serve as ground truth using 3D Slicer (https://www.slicer.org/).¹² For breast segmentation, the outer contour of each breast was traced. For FGT segmentation, manual thresholding was applied to classify the FGT voxels. All manual segmentations were reviewed and confirmed by a breast radiologist with seven years of experience.

All the segmentation masks with more information about the dataset used can be found at our website: https://sites.duke.edu/mazurowski/resources/breast-cancer-mri-dataset-fgt-and-breast-segmentation-2d/.

Our study was determined to be exempt from review by the Duke Health Institutional Review Board. The review board also determined that receiving informed consent is waived as we are using a publicly available database. All methods were performed accordance with the relevant guidelines and regulations.

2.4 Automated image segmentation using U-net

The U-net architecture has been successfully used in multiple medical imaging segmentation tasks.^13–18 This architecture consists of three sections: contraction, bottleneck, and expansion. The contraction section is made of many contraction blocks. Each block takes an input, applies two 3$\times$3 convolution layers, and then uses 2$\times$2 max pooling. The number of kernels or feature maps after each block doubles so that the architecture can learn the complex structures effectively. The bottommost layer mediates between the contraction layer and the expansion layer. It uses two 3$\times$3 CNN layers followed by 2$\times$2 up convolution layer.

Before training the network, we preprocessed the dataset in two steps. First, we normalized the image intensity between zero and one based on the 5th and 95th intensity percentiles. Next, all images were resized to 512$\times$512 using bilinear interpolation. ¹⁹

We selected 100 cases for development and 27 cases for testing. During the development process, 10% of the training cases (10 cases) were randomly chosen as the validation set. The models for breast and FGT segmentation were trained independently using the same U-net structure .The network was trained from random weights initialized using He’s initialization method.²⁰ The training processes included a total of 200 epochs before convergence. The Adam optimizer was used for back-propagation of gradients. To prevent overfitting, a term for L2 regularization was added to the binary cross-entropy loss. An initial learning rate of 1e-4 was used and decayed by 0.8 every 40 steps. The final segmented masks were predicted with a threshold of 0.5. These hyperparameters were chosen after iterative tests done on the validation set, with varying learning rates, batch sizes, and total epochs.

All of the code used for data preprocessing, model training, and model evaluation can be found online at: https://github.com/MaciejMazurowski/Breast-dense-tissue-segmentation

2.5 Evaluation of the model for segmentation accuracy

Performance of our automated segmentation model was assessed with the Dice similarity coefficient (DSC) which is a statistical validation metric that measures the similarity between two datasets and is a commonly used spatial overlap index to evaluate the performance of image segmentation models. The values of DSC range from 0 (indicating no spatial overlap between two sets of binary segmentation results) to 1 (indicating complete overlap). For this study, the DSCs were calculated by comparing the image segmentations produced by the model against the manual segmentations on the selected images to determine the accuracy of the model on the training and test sets.

2.6 Evaluation of the model for FGT percentage assessment

To evaluate the performance of the model in assessing the overall breast density, we used an extended testing set of all slices for 50 cases (including the 27 test set cases used in the previous sections). Three fellowship-trained and board-certified breast radiologists reviewed the pre-contrast images on these studies and classified the FGT into four categories according to the 2013 Breast Imaging-Reporting and Data System (BI-RADS) Atlas: almost entirely fat, scattered fibroglandular tissue, heterogeneous fibroglandular tissue, and extreme fibroglandular tissue (category a, b, c, and d, respectively).

The FGT percentage for the algorithm was computed based on the breast and FGT segmentations using the following formula²¹:

$$FGT percentage=\frac{\left|FGT\right|}{\left|Breast\right|} \times 100$$

The radiologist assessments were regarded as ground truth and were compared to the FGT percentage calculated by our algorithm. We assessed the association between the algorithm and the radiologists using Spearman’s correlation coefficient (r). The inter-reader variability among the three radiologists was measured with Cohen’s kappa coefficient (κ).

3.1 Segmentation performance of the model

The average DSC values of breast and FGT segmentations obtained on the validation and test sets are shown in Table 2. In the validation set, the DSC for breast and FGT segmentations were 0.870 and 0.744, respectively. In the testing dataset, the DSC for breast and FGT segmentations were 0.879 and 0.730, respectively. Figures 2 and 3 show examples of automated segmentation results versus ground truth from the test set for breast and FGT, respectively.

Table 2

DSC of breast and FGT segmentation
	Breast	FGT
Validation set	0.870	0.744
Test set	0.879	0.730

3.2 Correlation of predicted FGT percentage and BI-RADS breast density

The distribution of BI-RADS breast density assessments for the 50 cases (test set n = 27, extended test set n = 23) among the three radiologists was as follows: (a) almost entirely fat (16/150, 10.7%), (b) scattered fibroglandular tissue (52/150, 34.7%), (c) heterogeneous fibroglandular tissue (47/150, 31.3%), and (d) extreme fibroglandular tissue (35/150, 23.3%). Table 3 shows the confusion matrix for breast density assessments by different radiologists. The Cohen’s kappa coefficients were κ = 0.62 for radiologists A and B, κ = 0.38 for radiologists A and C, and κ = 0.57 for radiologists B and C. The correlation between the median BI-RADS breast density assessments and FGT percentage was r = 0.83 (p < 0.01). See Fig. 4 for the illustration of the relationship between individual radiologist assessments and FGT percentage.

Table 3

Confusion matrices comparing the BI-RADS breast density scores among the three radiologists. a = almost entirely fat; b = scattered fibroglandular tissue; c = heterogeneous fibroglandular tissue; d = extremely fibroglandular tissue.
		Radiologist A
		a	b	c	d
Radiologist B	a	3	0	0	0
	b	6	11	0	0
	c	0	7	9	1
	d	0	0	0	13

		Radiologist A
		a	b	c	d
Radiologist C	a	3	1	0	0
	b	6	9	2	0
	c	0	8	7	6
	d	0	0	0	8

		Radiologist B
		a	b	c	d
Radiologist C	a	1	3	0	0
	b	2	12	3	0
	c	0	2	14	5
	d	0	0	0	8

In this study, we developed a deep learning-based automated model to segment the breast and FGT as well as to calculate the FGT percentage. Our model provides a segmentation of FGT at a level similar to prior algorithms developed for this task.^21–28 (Table 4).

Table 4

Summary of breast MRI segmentation models with their associated Dice similarity coefficient values
Author (Year)	Technique	Test Performance
Did not apply deep learning
Wu et al. (2013)	Atlas-aided fuzzy C-means	DSC of 0.69 for reader 1 and 0.61 for reader 2
Ivanovska et al. (2014)	Level-set based model	DSC values of 0.96 and 0.83 for breast volume and FGT, respectively
Gubern-Merida et al. (2015)	Atlas-based probabilistic model	DSC values of 0.94 and 0.8 for breast volume and FGT, respectively
Razavi etl al. (2015)	Atlas-based probabilistic model + skin-fold removal	DSC value of 0.84 for FGT
Applied deep learning
Dalmis et al. (2017)	2 consecutive U-nets, 3 class U-net	DSC values: breast volume; FGT 2C U-nets: 0.94; 0.81 3C U-net: 0.93; 0.85 Atlas-based method: 0.86; 0.67 Sheetness-based method: 0.85
Ha et al. (2018)	3D CNN	DSC values of 0.95 and 0.81 for breast volume and FGT, respectively
Zhang et al. (2019)	U-net	DSC values of 0.86 and 0.83 for breast volume and FGT, respectively
Van der Velden et al. (2020)	3D regression CNN	Spearman’s correlation of p = 0.81 between estimated volumetric density and the ground truth volumetric density

There are two crucial elements of this study that set it apart from the prior work: (1) comparison of the model’s FGT percentage to the ground truth of radiologists, and (2) full transparency by providing data and the model for public use. Regarding the clinical significance, we correlated the FGT percentage calculated by our algorithm and the BI-RADS scores given by the radiologists. Previous studies have only compared the models’ performance to the manual segmentations done by radiologists to assess model performance but have not translated their findings into a scenario that simulates an actual clinical application setting. The key feature of an automated segmentation method is its potential to reduce subjectivity and reader variability in the qualitative assessments of breast density on MRI. Even though the breast density component has been a part of the BI-RADS scoring system since its initial inception, considerable inter- and intra-observer variability persists. The problem of variability has been shown in many mammography studies.^29–34 For example, Nicholson et al. reported a low inter-observer agreement of 49% with most of the breast density assessment agreement occurring at the two extremes (almost entirely fat vs extreme fibroglandular tissue).³¹ Kerlikowske et al. showed that the inter-reader and intra-reader agreements regarding breast density assessments were κ = 0.59 and κ = 0.72, respectively, indicating imperfect reliability of density interpretation even by the same reader.³⁴ Similar problems are present in the studies available in the literature that attempt to evaluate tissue density on MRI. For example, in a study to assess the interobserver variability of breast radiologists for the MRI BI-RADS lexicon, Grimm et al. reported k of 0.28 for inter-reader variability for background parenchymal enhancement.³⁵ Even in our study, the range of κ values was 0.38–0.62 indicating considerable variability among the three radiologists.

Another unique aspect of our study is that we provide full transparency of the data and code for a fully automated deep learning breast segmentation model that has been validated against board-certified radiologists. The manually segmented MR images from our heterogeneous dataset are now publicly available. To date, all of the studies involving FGT segmentation and breast density estimation have used private data for model development. One of the advantages of using an open-access database is that it allows for data transparency. Transparency is important in model training in that it facilitates model optimization and interpretation of the results. In addition to the benefits of the results transparency, an accessible large database promotes data reproducibility. The importance of using an external validation cohort has been emphasized in previous studies identifying radiomic features for breast disorders.^11,36,37 Similarly for model development, the public database will help promote its generalizability in the large, heterogeneous sample.

Our study had some limitations. While we annotated and shared the annotation masks, it still constitutes a relatively small number of slices. Furthermore, the distribution of types of scanners used to generate the MRI studies in our training set did not reflect the distribution of scanner types in the Duke Breast Cancer MRI dataset. These limitations can be minimized if our dataset is combined with others for model development. Since the images are publicly available, others could also contribute to creating a larger dataset by sharing their annotations for the publicly available images. A publicly available dataset will allow people to compare models on data from other sources to produce results that allow for a fairer comparison for model accuracy compared to our results in Table 3. This data set will also improve generalizability due to the heterogeneous nature of the cohort. Furthermore, due to our small sample size, our segmentation model should be validated in a larger cohort.

We conducted extensive annotation work and made the annotations publicly available. We also developed and tested a deep learning-based algorithm for segmentation of breast and fibroglandular tissue in breast MRIs. With further validation, our model can aid in providing a more objective evaluation of the breast density for clinical use in prognosticating patients’ likelihood of breast malignancy.

7 Author Contributions

M. A. M supervised the whole project. S. H. designed the study and created the machine learning model. S. H. and C. P. created the dataset and analyzed the results. C. O. L. reviewed the code and verified the results. L. J. G., J. A. B., and M. W. T., reviewed the dataset. All authors participated in preparing the manuscript.

8 Competing Interests

The authors declare no competing interests.

9 Data Availability Statement

All of the data and code used in this study is available publicly. The original data can be found at the Duke Breast Cancer MRI dataset online (https://doi.org/10.7937/TCIA.e3sv-re93). The annotation data is in the process of being made available publicly at The Cancer Imaging Archive. The code used in the study can be found at https://github.com/MaciejMazurowski/Breast-dense-tissue-segmentation.

10 Ethics Declarations

Momenimovahed, Z. & Salehiniya, H. Epidemiological characteristics of and risk factors for breast cancer in the world. Breast Cancer (Dove Med Press) 11, 151–164, doi:10.2147/bctt.S176070 (2019).
Carney, P. A. et al. Individual and combined effects of age, breast density, and hormone replacement therapy use on the accuracy of screening mammography. Ann Intern Med 138, 168–175, doi:10.7326/0003-4819-138-3-200302040-00008 (2003).
Freer, P. E. Mammographic breast density: impact on breast cancer risk and implications for screening. Radiographics 35, 302–315, doi:10.1148/rg.352140106 (2015).
Kolb, T. M., Lichy, J. & Newhouse, J. H. Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations. Radiology 225, 165–175, doi:10.1148/radiol.2251011667 (2002).
Advani, S. M. et al. Association of Breast Density With Breast Cancer Risk Among Women Aged 65 Years or Older by Age Group and Body Mass Index. JAMA Netw Open 4, e2122810, doi:10.1001/jamanetworkopen.2021.22810 (2021).
McCormack, V. A. & dos Santos Silva, I. Breast density and parenchymal patterns as markers of breast cancer risk: a meta-analysis. Cancer Epidemiol Biomarkers Prev 15, 1159–1169, doi:10.1158/1055-9965.Epi-06-0034 (2006).
Vachon, C. M. et al. Mammographic density, breast cancer risk and risk prediction. Breast Cancer Res 9, 217, doi:10.1186/bcr1829 (2007).
Niukkanen, A. et al. Quantitative Volumetric K-Means Cluster Segmentation of Fibroglandular Tissue and Skin in Breast MRI. J Digit Imaging 31, 425–434, doi:10.1007/s10278-017-0031-1 (2018).
Tagliafico, A. et al. Mammographic density estimation: Comparison among BI-RADS categories, a semi-automated software and a fully automated one. The Breast 18, 35–40, doi:https://doi.org/10.1016/j.breast.2008.09.005 (2009).
Alomaim, W. et al. Subjective Versus Quantitative Methods of Assessing Breast Density. Diagnostics (Basel) 10, doi:10.3390/diagnostics10050331 (2020).
Saha, A. et al. A machine learning approach to radiogenomics of breast cancer: a study of 922 subjects and 529 DCE-MRI features. Br J Cancer 119, 508–516, doi:10.1038/s41416-018-0185-8 (2018).
Fedorov, A. et al. 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 30, 1323–1341, doi:10.1016/j.mri.2012.05.001 (2012).
Badrinarayanan, V., Kendall, A. & Cipolla, R. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 2481–2495, doi:10.1109/TPAMI.2016.2644615 (2017).
Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K. & Yuille, A. L. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Trans Pattern Anal Mach Intell 40, 834–848, doi:10.1109/tpami.2017.2699184 (2018).
LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436–444, doi:10.1038/nature14539 (2015).
Zaharchuk, G., Gong, E., Wintermark, M., Rubin, D. & Langlotz, C. P. Deep Learning in Neuroradiology. American Journal of Neuroradiology, doi:10.3174/ajnr.A5543 (2018).
Krizhevsky, A., Sutskever, I. & Hinton, G. ImageNet Classification with Deep Convolutional Neural Networks. Neural Information Processing Systems 25, doi:10.1145/3065386 (2012).
Livne, M. et al. A U-Net Deep Learning Framework for High Performance Vessel Segmentation in Patients With Cerebrovascular Disease. Front Neurosci 13, 97, doi:10.3389/fnins.2019.00097 (2019).
Van der Velden, B. et al. Response monitoring of breast cancer on DCE-MRI using convolutional neural network-generated seed points and constrained volume growing. (2019).
He, K., Zhang, X., Ren, S. & Sun, J. in Proceedings of the 2015 IEEE International Conference on Computer Vision (ICCV) 1026–1034 (IEEE Computer Society, 2015).
Wu, S., Weinstein, S. P., Conant, E. F. & Kontos, D. Automated fibroglandular tissue segmentation and volumetric density estimation in breast MRI using an atlas-aided fuzzy C-means method. Med Phys 40, 122302, doi:10.1118/1.4829496 (2013).
Gubern-Mérida, A., Kallenberg, M., Mann, R. M., Martí, R. & Karssemeijer, N. Breast segmentation and density estimation in breast MRI: a fully automatic framework. IEEE J Biomed Health Inform 19, 349–357, doi:10.1109/jbhi.2014.2311163 (2015).
Ivanovska, T. et al. A level set based framework for quantitative evaluation of breast tissue density from MRI data. PLoS One 9, e112709, doi:10.1371/journal.pone.0112709 (2014).
Razavi, M. et al. in ICIAP.
Dalmış, M. U. et al. Using deep learning to segment breast and fibroglandular tissue in MRI volumes. Med Phys 44, 533–546, doi:10.1002/mp.12079 (2017).
Ha, R. et al. Fully Automated Convolutional Neural Network Method for Quantification of Breast MRI Fibroglandular Tissue and Background Parenchymal Enhancement. J Digit Imaging 32, 141–147, doi:10.1007/s10278-018-0114-7 (2019).
Zhang, Y. 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 26, 1526–1535, doi:10.1016/j.acra.2019.01.012 (2019).
van der Velden, B. H. M., Janse, M. H. A., Ragusi, M. A. A., Loo, C. E. & Gilhuijs, K. G. A. Volumetric breast density estimation on MRI using explainable deep learning regression. Sci Rep 10, 18095, doi:10.1038/s41598-020-75167-6 (2020).
Ciatto, S., Houssami, N. & Apruzzese, A. Categorizing breast mammographic density: intraand interobserver reproducibility of BI-RADS density categories. Breast 14, 269 (2005).
Eom, H. J. et al. Comparison of variability in breast density assessment by BI-RADS category according to the level of experience. Acta Radiol 59, 527–532, doi:10.1177/0284185117725369 (2018).
Nicholson, B. T. et al. Accuracy of assigned BI-RADS breast density category definitions. Acad Radiol 13, 1143–1149, doi:10.1016/j.acra.2006.06.005 (2006).
Ooms, E. A. et al. Mammography: Interobserver variability in breast density assessment. The Breast 16, 568–576, doi:https://doi.org/10.1016/j.breast.2007.04.007 (2007).
Berg, W. A., Campassi, C., Langenberg, P. & Sexton, M. J. Breast Imaging Reporting and Data System. American Journal of Roentgenology 174, 1769–1777, doi:10.2214/ajr.174.6.1741769 (2000).
Kerlikowske, K. et al. Variability and accuracy in mammographic interpretation using the American College of Radiology Breast Imaging Reporting and Data System. J Natl Cancer Inst 90, 1801–1809, doi:10.1093/jnci/90.23.1801 (1998).
Grimm, L. J. et al. Interobserver Variability Between Breast Imagers Using the Fifth Edition of the BI-RADS MRI Lexicon. American Journal of Roentgenology 204, 1120–1124, doi:10.2214/AJR.14.13047 (2015).
Fan, M. et al. Radiomic analysis reveals DCE-MRI features for prediction of molecular subtypes of breast cancer. PLoS One 12, e0171683, doi:10.1371/journal.pone.0171683 (2017).
Wu, J. et al. Identifying relations between imaging phenotypes and molecular subtypes of breast cancer: Model discovery and external validation. J Magn Reson Imaging 46, 1017–1027, doi:10.1002/jmri.25661 (2017).

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Fully Automated Deep Learning Method for Fibroglandular Tissue Segmentation in Breast MRI

Status:

Version 1

Abstract

Figures

1 Introduction

2 Material And Methods

2.1 Study population

2.2 Magnetic Resonance Imaging protocol

2.3 Image selection and annotation

2.4 Automated image segmentation using U-net

2.5 Evaluation of the model for segmentation accuracy

2.6 Evaluation of the model for FGT percentage assessment

3 Results

3.1 Segmentation performance of the model

3.2 Correlation of predicted FGT percentage and BI-RADS breast density

4 Discussion

5 Conclusion

Declarations

References

Additional Declarations

Status:

Version 1