Automatic Segmentation of Mammary Tissue using Computer

Simulations of Breast Phantoms and Deep-learning Techniques

Lucca R. Peregrino

, Jordy V. Gomes

, Tha

ıs G. do R

ego

, Yuri de A. M. Barbosa

1 a

Telmo de M. e Silva Filho

2 b

, Andrew D. A. Maidment

and Bruno Barufaldi

3,∗ c

Center of Informatics, Federal University of Para

ıba, Jo

ao Pessoa, Brazil

Department of Statistics, Federal University of Para

ıba, Jo

ao Pessoa, Brazil

Department of Radiology, University of Pennsylvania, 3640 Hamilton Walk, Philadelphia PA, U.S.A.

Keywords:

Phantoms, Tomosynthesis, Deep Learning, U-Net, Segmentation.

Abstract:

Digital breast tomosynthesis (DBT) has rapidly emerged for screening mammography to improve cancer de-

tection. Segmentation of dense tissue plays an important role in breast imaging applications to estimate cancer

risk. However, the current segmentation methods do not guarantee an ideal ground-truth in clinical practice.

Computer simulations provide ground-truth that enables the development of convolutional neural network

(CNN) applications designed for image segmentation. This study aims to train a CNN model to segment

dense tissue in DBT images simulated using anthropomorphic phantoms. The phantom images were simu-

lated based on clinical settings of a DBT system. A U-Net, a CNN model, was trained with 2,880 images using

a slice-wise approach. The U-Net performance was evaluated in terms of percent of density in the central slice

and volumetric breast density in the medio-lateral slices. Our results show that the U-Net can segment dense

tissue from DBT images with overall loss, accuracy, and intersection over union of 0.27, 0.93, and 0.62 in

the central slices, and 0.32, 0.92, and 0.54 in the medio-lateral slices, respectively. These preliminary results

allow us to explore the use of CNN architectures to segment dense tissue in clinical images, which is a highly

complex task in screening with DBT.

1 INTRODUCTION

Digital mammography (DM) and digital breast to-

mosynthesis (DBT) are considered the “gold stan-

dard” of care for breast cancer screening (Tice and

Feldman, 2008; Vedantham et al., 2015; Azar and El-

Said, 2013). These imaging modalities increase the

sensitivity in cancer detection and reduce the num-

ber of recall rates when compared to the traditional

screening with screen-ﬁlm (Vedantham et al., 2015).

Complementary tools such as computer-aided di-

agnosis systems and convolution neural network

(CNN) applications can facilitate the early cancer lo-

cation by enhancing and detecting lesions (Cheng

et al., 2006; Azar and El-Said, 2013), which poten-

tially improve the diagnosis on mammography exams.

Anthropomorphic breast phantoms have been

widely used for research and development of mam-

https://orcid.org/0000-0002-7779-0288

https://orcid.org/0000-0003-0826-6885

https://orcid.org/0000-0003-3954-3611

mography imaging systems (Caldwell and Yaffe,

1990; Carton et al., 2011). These phantoms simulate

the mammary tissue accurately in terms of size, vol-

ume, and composition. Simulations of breast phan-

toms can be used as data augmentation to support

CNN architectures (Lashgari et al., 2020), assisting

the lack of data from speciﬁc populations (Barufaldi

et al., 2018a). In addition, these simulations provide

ground-truth images (i.e., ideal reference), which is

not provided in clinical practice (Tunc¸ay and Akdu-

man, 2014).

Manual or semi-automatic methods have been

developed to obtain ground-truth from images by

segmenting and thresholding different ﬁndings (Rui

et al., 2018; Chatﬁeld et al., 2014; Valverde et al.,

2017). In medical imaging, these segmentation meth-

ods are commonly performed by experts in radi-

ology. Because of the subjectivity of inter-and/or

intra-readers, the output resulting from these meth-

ods may include variability and inaccuracy (Oliveira,

2017), while computer simulations provide the actual

ground-truth for the segmented image.

252

Peregrino, L., Gomes, J., G. do Rêgo, T., Barbosa, Y., Filho, T., Maidment, A. and Barufaldi, B.

Automatic Segmentation of Mammary Tissue using Computer Simulations of Breast Phantoms and Deep-learning Techniques.

DOI: 10.5220/0010310402520259

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 4: BIOSIGNALS, pages 252-259

ISBN: 978-989-758-490-9

The large variability of data with ground-truth

provided by computer simulations can improve the

performance of image segmentation with CNNs

(Hamarneh and Jassi, 2010; Li et al., 2009). The

ground-truth identiﬁes each breast tissue, structures,

and ﬁndings to be used as labels for the input images

required in CNNs (Hamarneh and Jassi, 2010).

This study aims to develop a CNN model for

tissue segmentation using computer simulations of

breast phantoms. The CNN model is trained and

tested using an U-Net architecture (Ronneberger

et al., 2015). Projections are simulated using the ac-

quisition geometry of a clinical DBT system. DBT

phantom images are reconstructed using a customized

increment between reconstructed slices (0.1 mm).

The U-Net architecture is trained using a slice-wise

approach to segment glandular tissue from 2,880 DBT

reconstructed images.

2 BACKGROUND

2.1 Antropomorphic Breast Phantoms

Anthropomorphic breast phantoms have been widely

used to conduct in-silico trials (Bakic et al., 2018;

Maidment, 2014; Abadi et al., 2020). Breast char-

acteristics such as shape, size, volume, and tissue

composition should be accurately simulated in ac-

cordance with the human anatomy (Tunc¸ay and Ak-

duman, 2014). In addition, breast anatomical fea-

tures (e.g., glandular segments, Cooper’s ligaments

and blood vessels) should be realistically simulated

(Elangovan et al., 2017).

To ensure that images simulated using anthropo-

morphic breast phantoms are comparable to clinical

mammograms, validation methods that rely on hu-

man visual inspection and/or computer analyses are

required. For example, Elangovan et al. (2017)

propose a method that can rapidly produce a mul-

tiplicity of different breast appearance models using

4-alternative forced choice (4-AFC). Using 4-AFCs,

they have shown that simulated and real images were

statistically indistinguishable by expert breast read-

ers.

However, the recruitment of breast experts needed

to validate anthropomorphic phantoms can be a chal-

lenging task, and in-silico trials have been designed

as attempt to simulate human readings. Badano et

al. (2018) and Bakic et al. have reproduced read-

ing interpretations reported on large scale clinical tri-

als designed for pre-market approval of novel imaging

technologies (Badano et al., 2018; Bakic et al., 2018).

These previous publications reported a successful use

of mathematical models to simulate virtual readers

and design virtual anthropomorphic phantoms.

In breast imaging, computer simulations usually

require the use of breast phantoms (Bakic et al., 2018;

Maidment, 2014; Abadi et al., 2020). In this study,

computer simulations of breast phantoms were used

to design a novel CNN application for image segmen-

tation using an U-Net architecture.

2.2 U-Net Architecture

CNNs are often used for image classiﬁcation and seg-

mentation (Rui et al., 2018; Chatﬁeld et al., 2014;

Valverde et al., 2017). The U-Net architecture is

a CNN that was developed for biomedical image

segmentation (Ronneberger et al., 2015). The ul-

timate goal of the U-Net architecture (Figure 1)

is the segmentation and localization of desired ob-

jects/structures highlighted in the input images (Paul,

2018). The major beneﬁt of the U-Net is that there

is no need to use a large number of images for train-

ing and testing (Ronneberger et al., 2015). Thus, U-

Net can be useful for segmentation of medical images,

due to the fact that the segmented ground-truth is not

available in clinical practice. Besides that, U-Net uses

a reduced amount of training parameters compared to

other CNN’s, such as SegNet (Badrinarayanan et al.,

2017).

The U-Net architecture consists of contraction

(encoder) and expansion (decoder) paths. A set of two

convolutions (3×3 kernel) and one maxpooling (2×2

kernel) with ReLU activation are performed in each

encoder layer. Similarly, each decoder layer starts

with an upsampling and a 2×2 convolution, followed

by two 3×3 convolutions with ReLU activation. The

encoder provides ﬁltered information (feature maps)

acquired during the contraction path to be interpreted

by the decoder. The decoder concatenates the out-

put of transposed convolution layers with the feature

maps acquired from the encoder at each layer. Finally,

an activation function is used to predict classes of the

input images based on previous knowledge (training

phase) obtained from ground-truth images (Academy,

2019).

The U-Net architecture has been used in sev-

eral medical applications (Norman et al., 2018; Sev-

astopolsky, 2017). For example, Tong et al. (2018)

developed an improved U-Net architecture to segment

pulmonary nodules from CT images. The authors

concluded that the accuracy of nodule segmentation

is comparable or superior to the manual segmentation.

A different U-Net application, developed by Norman

et al. (2018), has shown an improved segmentation of

cartilage and meniscus from knees using clinical MRI

Automatic Segmentation of Mammary Tissue using Computer Simulations of Breast Phantoms and Deep-learning Techniques

253

images. Similarly, the precision of the automatic seg-

mentation demonstrated to be comparable to the man-

ual segmentation of experts.

In breast imaging, Zhang et al. (2020) developed a

transfer learning application that uses U-Net and Seg-

Net architectures to segment whole-breasts from MRI

scans. The authors modiﬁed and adapted both archi-

tectures using slice-wise approaches and obtained av-

erage dice coefﬁcient results of 0.87 (independent test

data set). Although the authors presented compelling

results, they emphasized that recruiting experts was

challenging because of the limited and expensive time

from radiologists to manually segment breasts from

MRI scans. In addition, the authors could not obtain

the manual segmentation (ground-truth) of the entire

image dataset.

Several software have been developed to segment

breast glandular tissue and estimate volumetric breast

density or breast dense area, such as Volpara et al.

(2014) and LIBRA et al. (2012). These software

use image processing techniques (e.g., edge detection,

support-vector machine, etc.) to segment breast tissue

in clinical images, unlike the methods proposed in the

current study.

Our proposed method is based on computer sim-

ulations that do not require manual segmentation of

breast tissue.

3 MATERIALS AND METHODS

3.1 Computer Simulations

The OpenVCT framework (Barufaldi et al., 2018b)

was used to simulate anthropomorphic breast phan-

toms (Zhang et al., 2008; Pokrajac et al., 2012). The

breast phantoms are composed by voxel-materials (la-

bels) that represent various tissue types and air (Bar-

ufaldi et al., 2018a). The tissue types are simulated

using an octree-based recursive partitioning method

(Pokrajac et al., 2012). In this method, seed points

are randomly selected within the phantom interior and

used to simulate glandular and adipose tissue bounded

by ﬁbrous Cooper’s ligaments (Figure 2). These tis-

sue types are simulated to mimic the breast anatomy.

We combined and simulated all phantom param-

eters described in Table 1 (n=96). These param-

eters were selected based on previous publications

(Bakic et al., 2018; Barufaldi et al., 2019). Fi-

nally, a breast tissue compression was simulated using

a GPU-accelerated mesh software (Barufaldi et al.,

2018a). The compression was performed using a

medio-lateral (ML) view.

Table 1: Summary of the breast phantom parameters.

Anthropomorphic

Breast Phantoms

Parameters

Number of Phantoms (#) 96

Distribution of dense

compartments (%)

{1.0; 15.0;

25.0; 50.0}

Breast volume (mL) 700

Breast thickness (mm) 63.3

Voxel size (mm) {0.1; 0.2}

Number of

compartments (#)

{425; 850;

1275; 1700}

Compartment shape

{(0.1; 1.0; 1.0; 2.0),

(0.01; 1.0; 1.0; 4.0)}

Ligament thickness (mm) [0.1;0.18]

The 3D breast phantoms were “sliced” through

the entire volume in the sagittal orientation (Figure

2, left). In total, 633 2D slices (784×2053 pixels)

were acquired per phantom (ML view). The phan-

tom slices contain label maps at each 0.1 mm thick

of compressed breast (Figure 2, right). Each label

represents a different x-ray mass attenuation (Hubbell

and Seltzer, 1995) used to simulate DBT projections

(Feng and Sechopoulos, 2012). For each phantom, a

set of 15 x-ray projections was simulated using the

acquisition geometry of a clinical DBT system (Ta-

ble 2). The x-ray projections were simulated using a

GPU-enabled x-ray tracing algorithm (Siddon, 1985).

The exposure acquisition settings follow the auto-

matic exposure control from the DBT system (Feng

and Sechopoulos, 2012).

Table 2: Summary of the DBT acquisition parameters.

DBT System (model)

Selenia

Dimensions

X-Ray Imaging

Anode Material Tungsten

Filter Material Aluminum

Filter Thickness (mm) 0.7

Angular Range (

◦

) [±7.5,±15,±25]

Number of Projections (#) 15

Tube Motion Continuous

Detector

Detection Material a-Se

Detector Element Size(mm) 0.140 × 0.140

Number of Elements (#) 2048 × 1664

Detector Size (mm) 286.72 × 232.96

Source-Image Dis.(mm) 700.0

Rec. Voxel Size (mm) 0.1

A commercial reconstruction software (Briona

Std., Real-Time Tomography, Vilanova PA) was used

to reconstruct and to process each set of DBT pro-

jections (Chui et al., 2012). This software allows us

to reconstruct DBT images using customized recon-

struction voxel size. In this study, the DBT images

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

254

Figure 1: U-Net architecture used in this study. This architecture shows the ”u” structure resulting from the encoder and

decoder paths.

Figure 2: (Left) Volume view of compressed anthropomor-

phic breast phantom and (right) central slice (ML breast

view). The colormap represents the labels used to iden-

tify each voxel-material: (A) air, (B) skin, (C) adipose, (D)

glandular, (E-H) Cooper’s ligaments.

were reconstructed using 0.1 mm increments in depth.

In total, 633 reconstructed DBT images (1664×2048

pixels) were acquired per phantom.

3.2 Pre-processing Images

The reconstructed DBT images and correspondent la-

bel maps (ground-truth) were used as input for the

training and test stages of the U-Net architecture.

However, differences between the dimensions of the

input images will result in an ineffective CNN model.

Pre-processing techniques were required to match the

ground-truth to the respective reconstructed DBT im-

age.

The acquisition geometry (Table 2) was used to lo-

cate each label on the ground-truth and the respective

pixel value in the DBT image. Next, a cropping op-

eration was applied to each DBT image to eliminate

excessive background information from the DBT im-

ages (Figure 3, left and middle). Finally, the glandular

tissue label was thresholded and segmented from the

ground-truth to obtain binary masks used for training

the CNN model (Figure 3, right). After these pre-

processing steps, the reconstructed DBT image, label

map, and binary mask are matched (778×2,036 pix-

els). Finally, both input images, DBT reconstructed

image and binary mask, were normalized using the

maximum value in bits (2

and 2

, respectively), re-

sulting in images with pixel values in a [0, 1] interval.

The input images were resized to 512×512 pixels to

optimize the CNN model and reduce computational

burden.

3.3 Training the Model

To train the segmentation model, we modiﬁed the

original U-Net parameters (Zhixuhao, 2016; Ron-

neberger et al., 2015) using the programming lan-

guage Python. The optimization of the U-Net param-

eters, as well as training and testing were performed

Automatic Segmentation of Mammary Tissue using Computer Simulations of Breast Phantoms and Deep-learning Techniques

255

Figure 3: (Left) central slice of reconstructed DBT image,

(middle) label map, and (right) binary mask after matching

and cropping operations.

using a workstation equipped with Intel(R) Xeon(R)

CPU, 16GB RAM, and single graphics card NVIDIA

Quadro P5000.

In total, 45 pairs of images (reconstructed

DBT image and mask) acquired from 64 phantoms

(n=2,880 pairs) were used to train the U-Net model.

The image pairs were selected using 45 pairs of phan-

tom central slices, which contain regions with the

most amount of glandular tissue.

Our U-Net model (Figure 1) was trained using 120

epochs, batch size 4, and image size 512×512. The

weights of the training model were updated after each

iteration (n=720). These parameters were selected

and constrained based on memory used to train the

architecture models. The number of epochs was op-

timized based on loss and accuracy. It is important

to mention that the training models are saved every

epoch. The training model did not improve signiﬁ-

cantly after 120 epoch.

3.4 Evaluation Metrics

The binary crossentropy, accuracy, and intersection

over union were used to evaluate the performance of

the segmentation models. These metrics are deﬁned

as:

Binary Crossentropy (Loss) is a loss function

that calculates the difference between predicted labels

( ˆy) and true labels (y). The loss is computed following

Equation (1):

Loss(y, ˆy) = −(y · log(ˆy) + (1 − y) · log(1 − ˆy)) (1)

Accuracy (Acc) is a metric commonly used for

predictive models, calculating the proportion of cor-

rect predictions (CP) over the total instances. In our

approach, each pixel is an instance (Total). The for-

malism of accuracy is deﬁned in accordance with

Equation (2):

Acc =

Total

(2)

Intersection over Union (IoU) is a metric that

computes the segmented area ( ˆy) that corresponds to

the area of the mask (y), dividing what is in common

between them (intersection) by the whole (union).

Equation (3) shows the IoU calculation:

IoU =

area(ˆy)∩ area(y)

area(ˆy)∪ area(y)

(3)

Pearson’s correlation coefﬁcient (ρ) was calcu-

lated to evaluate the linear correlation between the

evaluated metrics and the percentage of glandular tis-

sue (PD%) in the phantom images. PD% is calculated

by using the ratio of glandular labels and non-air la-

bels (e.g., Figure 2).

The observed values were categorized by slice po-

sition through the phantom volume and volumetric

breast density (VBD).

4 EXPERIMENTAL ANALYSES

Two experiments were performed to test our segmen-

tation model. For both experiments, we used 32

unique breast phantoms. For each experiment, we

varied the slice location as input images. Similarly to

the training stage, only the central slice images were

used as input for the ﬁrst experiment. For the sec-

ond experiment, the entire phantom volume divided in

slices was used as input images. The phantom slices

and DBT reconstructed images close to the phantom

skin (about 1 cm in each extremity) were excluded

from the experimental analyses due to the lack of

glandular tissue for segmentation.

4.1 Using Central Slices

The experiment using only the central slices resulted

in a mean Loss, Acc and IoU of 0.27, 0.93, and 0.62,

respectively. Figures 4-6 show three examples from

our model segmentation using input images that con-

tain regions with different amounts of glandular tissue

(i.e., PD%). Note that the accuracy of the U-Net seg-

mentation varies with PD%.

The correlations between PD% and Loss, Acc,

and IoU were ρ=0.77, ρ=-0.85, and ρ=0.60, respec-

tively. This correlation analysis shows moderate

to high positive correlation between PD%, IoU and

Loss, and high negative correlation with Acc. That

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

256

Figure 4: (Left) central slice of reconstructed DBT image,

(middle) label map with glandular tissue highlighted in yel-

low (PD%=12%), and (right) binary segmentation. The seg-

mentation metrics for this input image were 0.18, 0.95, and

0.56 for Loss, Acc, and IoU, respectively.

Figure 5: (Left) central slice of reconstructed DBT image,

(middle) label map with glandular tissue highlighted in yel-

low (PD%=20%), and (right) binary segmentation. The seg-

mentation metrics for this input image were 0.19, 0.94, and

0.63 for Loss, Acc, and IoU, respectively.

said, these preliminary results indicate that the accu-

racy of the U-Net segmentation reduces signiﬁcantly

with PD% (p-value<0.001).

4.2 Using Medio-lateral Slices

We also evaluated the U-Net segmentation using all

phantom slices that contain glandular labels (ML

slices). In total, over 480 ML slices per phantom were

selected for this experiment. The U-Net segmentation

resulted in overall performance with mean Loss, Acc,

and IoU of 0.32, 0.92, and 0.54, respectively (Fig-

ure 7a). Note that there was a slight reduction in the

Acc and IoU metrics compared to the previous exper-

iment.

Figure 6: (Left) central slice of reconstructed DBT image,

(middle) label map with glandular tissue highlighted in yel-

low (PD%=34%), and (right) binary segmentation. The seg-

mentation metrics for this input image were 0.31, 0.91, and

0.70 for Loss, Acc, and IoU, respectively.

Figure 7b shows the results of the metrics catego-

rized by slice position through the entire phantom vol-

ume. The slice position also affects the U-Net perfor-

mance, since our CNN was trained using only central

slices. The relative difference in IoU between slice

positions can reach up to 20%. These differences can

also be seen in the Loss. However, these are prelimi-

nary results and a more detailed statistical analysis is

required to evaluate the U-Net performance in depth.

Finally, Figure 7c shows the Acc results catego-

rized by VBD. These boxplots show changes in U-

Net performance throughout the glandular volume of

breast phantoms. Note that the overall U-Net per-

formance tends to reduce with denser phantoms (i.e.,

higher VBD). Again, these are preliminary results and

a more detailed statistical analysis is required to sup-

port this observation.

5 CONCLUSIONS

These preliminary results show that our U-Net im-

plementation can segment glandular tissue from DBT

images with high accuracy. The computer simulations

are supervised, thus a known ground-truth is available

as input images for the U-Net training. Although our

results were based on simulations, our U-Net imple-

mentation can be potentially extended to clinical ap-

plications if a reasonable ground-truth data set is pro-

vided.

This U-Net application allows us to evaluate the

impact of 2D breast parameters (PD%) using particu-

lar slices and 3D breast parameters (VBD) using sets

of slices as input images. For future work, we will

provide a more complete statistical analysis of our

Automatic Segmentation of Mammary Tissue using Computer Simulations of Breast Phantoms and Deep-learning Techniques

257

(a)

(b)

(c)

Figure 7: Boxplots of metrics evaluated using reconstructed

slices obtained from 10.1-50.0 mm of phantom thickness

(0.1 mm increment). (a) Observed values categorized by

metrics, (b) as a function of slice position, and (c) as a func-

tion of VBD. Note that there is a difference in y-scale in (c),

compared to (a) and (b).

dataset and further explore the use of 3D CNNs for

volume segmentation.

ACKNOWLEDGEMENTS

Funding for the research is provided by the following

grants: BWF IRSA 1016451, DoD W81XWH-18-1-

0082, and AAPM 2020 Research Seed Grant.

REFERENCES

Abadi, E., Segars, W. P., Tsui, B. M., Kinahan, P. E.,

Bottenus, N., Frangi, A. F., Maidment, A., Lo, J.,

and Samei, E. (2020). Virtual clinical trials in med-

ical imaging: a review. Journal of Medical Imaging,

7(4):042805.

Academy, D. S. (2019). Deep learning book.

Azar, A. T. and El-Said, S. A. (2013). Probabilistic neural

network for breast cancer classiﬁcation. Neural Com-

puting and Applications, 23(6):1737–1751.

Badano, A., Graff, C. G., Badal, A., Sharma, D., Zeng,

R., Samuelson, F. W., Glick, S. J., and Myers, K. J.

(2018). Evaluation of digital breast tomosynthesis

as replacement of full-ﬁeld digital mammography us-

ing an in silico imaging trial. JAMA network open,

1(7):e185474–e185474.

Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017).

Segnet: A deep convolutional encoder-decoder ar-

chitecture for image segmentation. IEEE transac-

tions on pattern analysis and machine intelligence,

39(12):2481–2495.

Bakic, P. R., Barufaldi, B., Higginbotham, D., Weinstein,

S. P., Avanaki, A. N., Espig, K. S., Xthona, A., Kimpe,

T. R. L., and Maidment, A. D. A. (2018). Virtual

clinical trial of lesion detection in digital mammog-

raphy and digital breast tomosynthesis. In Lo, J. Y.,

Schmidt, T. G., and Chen, G.-H., editors, Medical

Imaging 2018: Physics of Medical Imaging, volume

10573, pages 30 – 42. International Society for Optics

and Photonics, SPIE.

Barufaldi, B., Bakic, P., and Maidment, A. (2019).

Multiple-reader, multiple-case ROC analysis for de-

termining the limit of calciﬁcation detection in to-

mosynthesis . In Schmidt, T. G., Chen, G.-H., and

Bosmans, H., editors, Medical Imaging 2019: Physics

of Medical Imaging, volume 10948, pages 157 – 163.

International Society for Optics and Photonics, SPIE.

Barufaldi, B., Bakic, P. R., Pokrajac, D. D., Lago, M. A.,

and Maidment, A. D. (2018a). Developing popula-

tions of software breast phantoms for virtual clinical

trials. In 14th International Workshop on Breast Imag-

ing (IWBI 2018), volume 10718, page 107181U. Inter-

national Society for Optics and Photonics.

Barufaldi, B., Higginbotham, D., Bakic, P. R., and Maid-

ment, A. D. A. (2018b). OpenVCT: a GPU-

accelerated virtual clinical trial pipeline for mammog-

raphy and digital breast tomosynthesis. In Lo, J. Y.,

BIOSIGNALS 2021 - 14th International Conference on Bio-inspired Systems and Signal Processing

258

Schmidt, T. G., and Chen, G.-H., editors, Medical

Imaging 2018: Physics of Medical Imaging, volume

10573, pages 1333 – 1340. International Society for

Optics and Photonics, SPIE.

Caldwell, C. B. and Yaffe, M. J. (1990). Development of an

anthropomorphic breast phantom. Medical physics,

17(2):273–280.

Carton, A.-K., Bakic, P., Ullberg, C., Derand, H., and Maid-

ment, A. D. (2011). Development of a physical 3d

anthropomorphic breast phantom. Medical physics,

38(2):891–896.

Chatﬁeld, K., Simonyan, K., Vedaldi, A., and Zisserman,

A. (2014). Return of the devil in the details: Delving

deep into convolutional nets. CoRR, abs/1405.3531.

Cheng, H.-D., Shi, X., Min, R., Hu, L., Cai, X., and Du, H.

(2006). Approaches for automated detection and clas-

siﬁcation of masses in mammograms. Pattern recog-

nition, 39(4):646–668.

Chui, J. H., Pokrajac, D. D., Maidment, A. D. A., and Ba-

kic, P. R. (2012). Roadmap for efﬁcient paralleliza-

tion of breast anatomy simulation. In Pelc, N. J.,

Nishikawa, R. M., and Whiting, B. R., editors, Med-

ical Imaging 2012: Physics of Medical Imaging, vol-

ume 8313, pages 1369 – 1378. International Society

for Optics and Photonics, SPIE.

Elangovan, P., Mackenzie, A., Dance, D. R., Young, K. C.,

Cooke, V., Wilkinson, L., Given-Wilson, R. M., Wal-

lis, M. G., and Wells, K. (2017). Design and validation

of realistic breast models for use in multiple alterna-

tive forced choice virtual clinical trials. Physics in

Medicine & Biology, 62(7):2778.

Feng, S. S. J. and Sechopoulos, I. (2012). Clinical digital

breast tomosynthesis system: dosimetric characteriza-

tion. Radiology, 263(1):35–42.

Hamarneh, G. and Jassi, P. (2010). Vascusynth: Simu-

lating vascular trees for generating volumetric image

data with ground-truth segmentation and tree anal-

ysis. Computerized medical imaging and graphics,

34(8):605–616.

Hubbell, J. H. and Seltzer, S. M. (1995). Tables of x-

ray mass attenuation coefﬁcients and mass energy-

absorption coefﬁcients 1 kev to 20 mev for elements

z= 1 to 92 and 48 additional substances of dosimetric

interest. Technical report, National Inst. of Standards

and Technology-PL, Gaithersburg, MD (United . . . .

Lashgari, E., Liang, D., and Maoz, U. (2020). Data aug-

mentation for deep-learning-based electroencephalog-

raphy. Journal of Neuroscience Methods, page

108885.

Li, C. M., Segars, W. P., Tourassi, G. D., Boone, J. M., and

Dobbins III, J. T. (2009). Methodology for generat-

ing a 3d computerized breast phantom from empirical

data. Medical physics, 36(7):3122–3131.

Maidment, A. D. (2014). Virtual clinical trials for the as-

sessment of novel breast screening modalities. In

International Workshop on Digital Mammography,

pages 1–8. Springer.

Norman, B., Pedoia, V., and Majumdar, S. (2018). Use of

2d u-net convolutional neural networks for automated

cartilage and meniscus segmentation of knee mr imag-

ing data to determine relaxometry and morphometry.

Radiology, 288(1):177–185.

Oliveira, W. d. S. (2017). Consenso de segmentac¸

oes de

imagens usando classiﬁcac¸

ao de padr

oes. PhD thesis,

Universidade Federal de Pernambuco.

Paul, S. (2018). Learn how to train u-net on your

dataset. https://medium.com/coinmonks/learn-how-

to-train-u-net-on-your-dataset-8e3f89fbd623.

Pokrajac, D. D., Maidment, A. D., and Bakic, P. R. (2012).

Optimized generation of high resolution breast an-

thropomorphic software phantoms. Medical physics,

39(4):2290–2302.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net:

Convolutional networks for biomedical image seg-

mentation. In International Conference on Medical

image computing and computer-assisted intervention,

pages 234–241. Springer.

Rui, T., Zou, J., Zhou, Y., Fei, J., and Yang, C. (2018).

Convolutional neural network feature maps selection

based on lda. Multimedia Tools and Applications,

77(9):10635–10649.

Sevastopolsky, A. (2017). Optic disc and cup segmentation

methods for glaucoma detection with modiﬁcation of

u-net convolutional neural network. Pattern Recogni-

tion and Image Analysis, 27(3):618–624.

Siddon, R. L. (1985). Fast calculation of the exact radio-

logical path for a three-dimensional ct array. Medical

physics, 12(2):252–255.

Tice, J. and Feldman, M. (2008). Full-ﬁeld digital mam-

mography compared with screen-ﬁlm mammography

in the detection of breast cancer: Rays of light through

dmist or more fog? Breast cancer research and treat-

ment, 107:157–65.

Tunc¸ay, A. H. and Akduman, I. (2014). Realistic microwave

breast models through t1-weighted 3-d mri data. IEEE

Transactions on Biomedical Engineering, 62(2):688–

698.

Valverde, S., Cabezas, M., Roura, E., Gonz

alez-Vill

S., Pareto, D., Vilanova, J. C., Rami

o-Torrent

a, L.,

Rovira, A., Oliver, A., and Llad

o, X. (2017). Improv-

ing automated multiple sclerosis lesion segmentation

with a cascaded 3d convolutional neural network ap-

proach. CoRR, abs/1702.04869.

Vedantham, S., Karellas, A., Vijayaraghavan, G. R., and

Kopans, D. B. (2015). Digital breast tomosynthesis:

State of the art. Radiology, 277(3):663–684. PMID:

26599926.

Zhang, C., Bakic, P., and Maidment, A. (2008). Develop-

ment of an anthropomorphic breast software phantom

based on region growing algorithm - art. no. 69180v.

Proc SPIE, 6918.

Zhixuhao (2016). Implementation of deep

learning framework – unet, using keras.

https://github.com/zhixuhao/unet.

Automatic Segmentation of Mammary Tissue using Computer Simulations of Breast Phantoms and Deep-learning Techniques

259