Detection of Microcalcifications in Digital Breast Tomosynthesis

using Faster R-CNN and 3D Volume Rendering

Ana M. Mota

, Matthew J. Clarkson

, Pedro Almeida

and Nuno Matela

Instituto de Biofísica e Engenharia Biomédica, Faculdade de Ciências da Universidade de Lisboa, Lisboa, Portugal

Department of Medical Physics and Biomedical Engineering and the Centre for Medical Image Computing (CMIC),

University College London, London, U.K.

Keywords: Digital Breast Tomosynthesis, Faster R-CNN, Volume Rendering, Microcalcification Clusters.

Abstract: Microcalcification clusters (MCs) are one of the most important biomarkers for breast cancer and Digital

Breast Tomosynthesis (DBT) has consolidated its role in breast cancer imaging. As there are mixed

observations about MCs detection using DBT, it is important to develop tools that improve this task.

Furthermore, the visualization mode of MCs is also crucial, as their diagnosis is associated with their 3D

morphology. In this work, DBT data from a public database were used to train a faster region-based

convolutional neural network (R-CNN) to locate MCs in entire DBT. Additionally, the detected MCs were

further analyzed through standard 2D visualization and 3D volume rendering (VR) specifically developed for

DBT data. For MCs detection, the sensitivity of our Faster R-CNN was 60% with 4 false positives. These

preliminary results are very promising and can be further improved. On the other hand, the 3D VR

visualization provided important information, with higher quality and discernment of the detected MCs. The

developed pipeline may help radiologists since (1) it indicates specific breast regions with possible lesions

that deserve additional attention and (2) as the rendering of the MCs is similar to a segmentation, a detailed

complementary analysis of their 3D morphology is possible.

1 INTRODUCTION

Breast cancer is the type of cancer with higher

incidence, among all cancers and both sexes, and it

still represents the biggest cause of cancer mortality

among women (Sung et al., 2021). The mortality rate

from this disease has been decreasing is the last

decades due to the new therapies and the

implementation of screening programs for early

detection (Tabár et al., 2019).

The use of Digital Breast Tomosynthesis (DBT)

has been confirming its potential to address the tissue

overlapping limitations of Digital Mammography

(DM), the gold standard for breast screening until

recently. In fact, by including synthetic

mammographies generated from DBT data, DBT

alone is now used as a stand-alone modality to replace

DM (Bernardi et al., 2016; Food and Drug

https://orcid.org/0000-0002-1931-294X

https://orcid.org/0000-0002-5565-1252

https://orcid.org/0000-0001-5247-4011

https://orcid.org/0000-0002-8048-7896

Administration (FDA) U.S. , 2013; Freer et al., 2017;

Gilbert et al., 2015; Hofvind et al., 2018; Lång et al.,

2016; Zackrisson et al., 2018). DBT volume data can

be analyzed in depth through several 2D slices

(standard visualization slice-by-slice). This multi-

slice inspection leads to a longer analysis time

(because instead of two images, radiologists have to

inspect an average of sixty images per patient), which

represent a problem in daily practice and screening

environment (Caumo et al., 2018; Good et al., 2008;

Gur et al., 2009).

Computer-Aided Detection (CAD) systems based

on DBT have been implemented and evaluated in an

attempt to shorten the reading time while maintaining

the radiologist performance. However, despite the

efforts and improvements already achieved, due to

the high false positive (FP) rates and low specificity,

these CAD systems have not reached a level of

Mota, A., Clarkson, M., Almeida, P. and Matela, N.

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering.

DOI: 10.5220/0010938800003123

In Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2022) - Volume 2: BIOIMAGING, pages 80-89

ISBN: 978-989-758-552-4; ISSN: 2184-4305

performance that can be translated into a true

improvement in the real screening of breast cancer

(Fenton et al., 2007; Katzen & Dodelzon, 2018;

Lehman et al., 2015; Sechopoulos, Teuwen, & Mann,

2020).

On the other hand, a different type of

visualization, such as 3D volume rendering (VR),

may play an important complementary role in breast

cancer diagnosis (Venson et al., 2017). With a

visualization of the object through multiple angles,

one of the advantages of VR is to provide an intuitive

understanding of the underlying data at once. In

addition, as VR yields a true depth perception

(Suetens, 2009), it can help in the analysis of lesions

such as microcalcification clusters (MCs), sometimes

referred as harder to detect in DBT. These MCs are

often spread across several slices in the slice-by-slice

visualization, making the interpretation difficult. In

this way, a better understanding of its true 3D

morphology is important to differentiate between

benign and malignant microcalcifications.

In recent years, the increase in computational

power and bigger datasets have allowed the

development of algorithms for automatic object

detection with deep learning. The region-based

convolutional neural networks (R-CNNs) are one of

the main current focuses of research and development

of these methods (Girshick, Donahue, Darrell, &

Malik, 2014). As R-CNN and its descendent “fast R-

CNN” (Girshick, 2015) are computationally

expensive and extremely slow, another method has

emerged: “Faster R-CNN” (Ren, He, Girshick, &

Sun, 2015). With this object detection network, both

the CNN-based regional proposals and the regional

classification module are trained together with

significant weight sharing, led to increased sensitivity

for object detection and faster speed.

The published studies that use deep CNNs to

detect and localize lesions in DBT are still very

limited. In fact, the few works that exist are related

with the detection of soft tissue lesions (Buda et al.,

2020; Fotin, Yin, Haldankar, Hoffmeister, &

Periaswamy, 2016; Lai, Yang, & Li, 2020; Samala et

al., 2016). Regarding the use of Faster R-CNN in

particular, (Fan et al., 2019) developed a CAD system

for masses detection in DBT using a Faster R-CNN,

which is later compared to a framework of a 3D-Mask

R-CNN for mass detection and segmentation (Fan et

al., 2020). (Li et al., 2021) propose a Faster R-CNN

that uses mammary gland distribution as a prior

information to detect architectural distortions in DBT.

In this paper, a Faster R-CNN was trained for

detecting MCs in DBT. The aim is to input a whole

DBT image into the network and have a direct answer

about the localization or absence of MCs. This

information about the location is then introduced into

a 3D VR visualization software so that a 3D volume

of interest containing the predicted MCs can be

obtained. A public simulated database was used and

the preliminary results obtained are presented. To the

best of our knowledge, this is the first study of

automatic localization of MCs in whole DBT images

and the first time the DBT output of a deep CNN is

rendered and presented as a 3D volume of interest.

2 MATERIALS AND METHODS

This work was implemented on the MATLAB

R2020a and a NVIDIA Quadro P4000 GPU computer

was used.

2.1 Database and Pre-processing

The public database of Virtual Imaging Clinical Trial

for Regulatory Evaluation (VICTRE) project which

contains a total of 2986 virtual realistic patients

imaged with DBT was used (Badano et al., 2018;

VICTRE, 2018). This database contains cases

without lesions (absent) and with malignant masses

and MCs. For training, only cases with MCs were

considered (915 in total: 665 complete breast images

and 250 images containing only MCs) and for the

testing, absent and MC cases were included (280 and

284 complete breast images, respectively). Each case

with lesion contains four MCs consisting of 5

calcified lesions modelled as 195, 179, and 171 μm of

solid calcium oxalate.

In addition to the information about the presence

or absence of MCs, in cases where MCs were present,

information about the corresponding bounding boxes

(BBs) was also given to the network. This

information, in the form of x, y and z coordinates as

well as width and height, is s in the VICTRE database.

We adopted the usual distribution of breast

density in the general population: 10% fatty, 40%

scattered, 40% heterogeneous and 10% dense. The

reconstructed cases have different dimension in x, y

and z, depending on breast density: 1624 × 1324 ×

62 , 1421 × 1024 × 57 , 1148 × 753 × 47 and

1130 × 477 × 38 for fatty, scattered, heterogeneous

and dense breasts, respectively, with a voxel size of

0.085 × 0.085 × 1 𝑚𝑚



The data intensity was first normalized between 0

and 1 and then squared to highlight the higher

intensity values belonging to the MCs, while

attenuate the lower ones. With this pre-processing

step our aim was to specifically increase the contrast

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering

of regions of higher intensities. In addition, through

binarization and region growing operations, binary

masks that keep information belonging to the breast

and make everything else zero were created

(background suppression).

2.2 Faster R-CNN Object Detector

Faster R-CNN is based on a CNN and a region

proposal network (RPN) for detecting, localizing and

classifying objects in an image. The CNN module

(typically a pre-trained CNN), outputs a set of feature

maps and, for that reason, it is also called feature

extraction network. In our work, we used the ResNet-

18 model, trained on more than a million images from

the ImageNet database ("ImageNet," 2021). The RPN

is on top of the last convolutional layer of the CNN

and it uses default bounding boxes (anchors) with

different sizes and aspect ratios over the feature maps

generated from pre-trained CNN in order to find

objects with varying sizes and shapes. It is trained to

output a set of object proposals on the image, each

with an “objectness” score, regardless of the class of

the object (it only looks if it is an object or

background). The boxes with the highest score are

called region proposals and are introduced in another

branch of the network were they are resampled to a

fixed size (ROI Pooling) and, typically using few

fully connected layers, the class of the object present

in the boundary boxes is determined. Further details

about Faster R-CNN can be found in the original

paper (Ren et al., 2015). The main parameters used to

define our Faster R-CNN are presented in Table 1.

Table 1: Parameters used to design the Faster R-CNN.

Input size 224x224x3

Anchor Boxes 42x27; 63x45; 45x41

Pre-trained CNN ResNet-18

2.3 Faster R-CNN Training

The Faster R-CNN was trained using the end-to-end

method, where the RPN and the region classification

networks were trained simultaneously along 660k

iterations. Table 2 presents the main training options

defined for this work.

During training, several regions of the image are

processed from the training database. The positive

and negative overlap range properties control which

image regions are used for training. This overlap ratio

is defined as the Intersection over Union (IoU) metric

that describes the extent of overlap between two

boxes (ground truth and predicted BB). The greater

the region of overlap, the greater the IOU. The model

was trained to minimize the mean square error loss

between the predicted BBs and the ground truth using

the Stochastic Gradient Descent optimizer

(MathWorks, 2021).

Table 2: Options used to train the Faster R-CNN.

Solver Stochastic Gradient

Descent w momentu

Momentu

0.9

Size of mini-

atch 1

Learnin

rate 1e-3

Factor for L

ularization 5e-4

Trainin

metho

-to-en

Positive Overla

Ran

e [0.3 1]

ative Overla

Ran

e [0 0.1]

To prevent overfitting, each image in the training

set was augmented by random reflection in the left-

right direction and rotation between -20º and 20º. In

addition, a L

regularization term for the weight decay

was introduced in the loss function.

2.4 Evaluation Metrics

The network's ability to accurately detect and locate

the MCs was evaluated through the Free-response

Receiver Operating Characteristic (FROC) curve

(Bunch, Hamilton, Sanderson, & Simmons, 1977).

To obtain a point on the FROC curve, a threshold

value is fixed and only the findings that have scores

above that threshold are selected. Then the sensitivity

(true positive fraction) and mean number of FPs per

image are determined.

2.5 Data Visualization

Figure 1 shows the scheme followed during and after

Faster R-CNN training. A testing set is evaluated for

the detection of MCs using the trained Faster R-CNN

and the output results (predicted BBs) are visualized.

In addition to the standard 2D visualization, the output

detection was also analyzed through 3D visualization

with VR. The 2D visualization was performed by

calculating the 2D maximum intensity projection

(MIP) considering the slice where the cluster was

detected and the four adjacent slices (two down and

two up). The 3D visualization was performed through

VR with 3D MIP considering the same slices.

The Visualization Toolkit library (VTK) version

7.1.0. (Kitware, New York, EUA) (Schroeder,

Martin, & Lorensen, 2006; VTK, 2020) was used to

develop 3D specific software in order to visualize

DBT data through VR. The opacity/color transfer

BIOIMAGING 2022 - 9th International Conference on Bioimaging

Figure 1: Pipeline followed for connection between the output of the trained Faster R-CNN and the volume rendering

visualization of the detected object.

functions for an adequate rendering of these data were

calculated accordingly to previous work (Mota,

Clarkson, Orvalho, Almeida, & Matela, 2020).

In VR, changing the azimuth of a camera rotates

its position around the focal point (Schroeder et al.,

2006) allowing an immediate notion of the entire

volume in 3D. In this way, the volume of interest

containing the detected MCs is presented from

several angles (from 0º to 90º).

3 RESULTS

The training of 660k iterations was performed during

12 days. The analysis of one test image was done in

0.6 seconds (mean time) and for an entire DBT

volume our Faster R-CNN needed, on average, 29

seconds (depending on the size).

3.1 Faster R-CNN Detection

Figure 2 presents the FROC curve for the performance

of the training model to accurately detect and locate the

MCs for several thresholds. In addition, the

discriminative sensitivity values obtained for less than

8 FP /image are detailed in the Table 3.

3.2 Data Visualization

Four examples of detection output, including the FPs

(yellow) and true positives (green) BBs, obtained

with a threshold of 0.9 are presented in Figure 3. The

corresponding score is also shown. As described,

each detected MC is presented through two

visualization modes: 2D slice-by-slice and 3D VR.

As 3D VR is inspected through several angles (0,

22.5º, 45º, 67.5º and 90º), 2D MIP slice-by-slice is

presented using xy and xz representations for

comparison with VR 0º and 90º, respectively.

Figure 2: The FROC curve for the test dataset.

Table 3: The sensitivity values for less than 8 FP/image.

Sensitivity

(%)

FP/image

# MCs

detected

# MCs

undetected

40 0.1 125 159

47 0.2 146 138

51 0.8 158 126

54 1.8 170 114

57 2.7 178 106

59 3.2 184 100

61 4.8 186 98

62 5.7 194 90

66 7.8 206 78

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering

Figure 3: Example of four detection outputs obtained with a threshold of 0.9. Green: Ground truth BB; Yellow: predicted BB

(without score: FPs, with score: true positives). The predicted results are visualized with 2D slice-by-slice represented through

xy and xz planes and 3D VR with five different angles (0º, 22.5º, 45º, 67.5º and 90º).

BIOIMAGING 2022 - 9th International Conference on Bioimaging

In Figure 4 (a) are presented four examples of

missed detections (false negatives) and on Figure 4

(b) four incorrect detections (FPs). The detection

results are then visualized through 2D slice-by-slice

and 3D VR at xy and 0º, respectively.

Visualization with 3D VR is very flexible and

includes parameters that can significantly change its

appearance, as is the case of transfer functions. The

Figure 5 shows the displays of four detected MC

obtained with 2D visualization and 3D VR using two

different transfer functions.

4 DISCUSSION AND

CONCLUSIONS

In this work, a Faster R-CNN detector was trained to

detect MCs in DBT data and the preliminary results

obtained were analyzed through two different forms

of visualization: standard 2D slice-by-slice and 3D

VR specifically developed for DBT. VR is presented

as a supplementary visualization of the detected MCs,

providing a more detailed and high quality

complementary information.

A DBT dataset from the publicly available

database at The Cancer Imaging Archive website

(VICTRE, 2018) was used. The train dataset

consisted in entire DBT images and also some regions

of interest containing only the MCs. These smaller

regions were included because the DBT images are

much bigger than the ground truth boxes of MCs,

reaching ratios of 30:1. As the size of the images was

not changed in order keep the necessary spatial

resolution to see the small microcalcifications, it was

important to have training inputs with an emphasis on

the object to be detected. Nevertheless, the test

dataset only contains entire images, as happens in

clinical or screening practice.

In this type of lesion detection task, the time

required for the detector to give an answer about the

input data is very important because it should be

useful in real time clinical practice. 29 seconds to

analyze a volume of DBT data (which can comprise

~130 million voxels) is reasonable but this value can

be improved using computers with greater power.

Also, this time is highly influenced by the feature

extraction network. For this reason, in this

preliminary work, we chose a network with a

reasonable balance between time and accuracy

(ResNet-18). However, other pre-trained networks

that may show better results and different detection

times should be studied.

The most used metric to analyze the performance

of this type of detector is the FROC curve. The results

obtained with this curve in Figure 2 and Table 3 reveal

that it was possible to achieve a sensitivity of around

60% with 4 FP/image. These preliminary results are

promising but need further improvement by adding

more training data, optimizing some network parame-

ters, training over a greater number of iterations and, as

already mentioned, using different pre-trained CNNs.

Figure 4: Example of four missed detections (false negatives) and four incorrect detections (FPs). The BB are visualized

thourgh 2D slice-by-slice in xy and 3D VR ar 0º.

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering

Figure 5: Example of four detection outputs. Green: Ground truth BB; yellow: predicted BB. The predicted results are

visualized with 2D MIP slice-by-slice represented through xy and xz planes and 3D VR with two different angles (0º and 90º).

Each 3D VR was obtained using two different transfer functions, allowing different levels of MC segmentations.

BIOIMAGING 2022 - 9th International Conference on Bioimaging

The output results of Figure 3 were obtained by

using a threshold of 0.9 (i.e., only scores above 0.9

were considered), which corresponds to a sensitivity

of about 50% for 0.8 FP/image. Four examples of

output from the Faster R-CNN were presented. The

number of FP found for this threshold (yellow BBs

without a score) varies from three (in the first case) to

zero (in the third and fourth case). The correctly

detected MCs by the yellow BBs with scores that

overlap the true BBs (green) were observed using the

two mentioned visualization methods. In general, the

MCs have a reasonable visibility in the xy plane with

the 2D MIP slice-by-slice over five adjacent slices but

are distorted in xz, losing some definition due to the

larger voxel size in z. The 3D VR at 0º and 90º can be

directly compared with the 2D visualization in the xy

and xz planes, respectively. For all cases, there is

better contrast and less noise in the VR at 0º, with

better discernment of the MCs. This superior

definition is noticeable when comparing the VR at 90º

with the xz plane of the 2D visualization. In the VR at

90º there is a clear discrimination of the MCs, and it

is possible to observe quite clearly the calcifications

individually and with some degree of reality.

It is also important to analyze some situations

where the detection was not correct (Figure 4). In the

case of false negatives, there were prominent lesions

that the algorithm did not detect (Figure 4 (a) last

column) and others where the MCs were somehow

masked, making their detection difficult (Figure 4 (a)

third column). In the case of FPs, in fact, there were

some situations where, even to the human eye, doubt

could be raised (Figure 4 (b) second and third

column). But, in the remaining situations, there is

essentially a spiculated noise that was interpreted as

MC. It is therefore important to further improve the

quality of detection.

On the other hand, the flexibility of visualization

using VR is demonstrated with the images in Figure

5. In addition to have the spatial distribution in the

three directions (x, y and z), with different transfer

functions we can filter the data to a greater or lesser

extent and, thus, segment better some lesions, such as

MCs. The transfer functions used in this work have

the opacity/color on the y-axis and the intensity

values on the x-axis. For intensities below a "A" value

the object data is transparent, while intensity values

above "B" (A<B) correspond to completely opaque

voxels. Between A and B the opacity values follow a

linear distribution. From transfer function 1 to

transfer function 2 (Figure 5) the value of B has been

increased to reduce the contribution to the

visualization of objects with lower intensities, making

those with higher intensities stand out, such as MCs.

In his way, it was possible to obtain a "cleaner"

visualization, as seen in Figure 5 in column of the

transfer function 2. This rendering parameter is a

great advantage in noisy data as can be seen in the last

case of Figure 5.

During training, no distinction was made between

the different types of breast density. However,

different densities correspond to data with slightly

different histograms. In the detection/analysis step, it

is important to understand if the detector behaves in

the same way for different densities (for example, it

is known that some lesions are more difficult to detect

in dense breasts than in fat breasts). From the

comparison made between the detection and

visualization of the four density groups, we can infer

that there were no differences between them.

As already mentioned, as far as we know, this is

the first work about MCs detection and localization in

a whole DBT image using deep learning CNNs such

as Faster R-CNN. Of the few published works found

in this area, all refer to soft tissue as masses. (Fan et

al., 2019) developed a CAD system for the

prescreening of ROIs and discrimination of true

masses and FPs in DBT using a Faster R-CNN. For

lesion-based mass detection, the sensitivity of their R-

CNN based CAD was 90% at 1.54 FP/volume. Later,

the same group, compared this work to a framework

of a 3D-Mask R-CNN for mass detection and

segmentation (Fan et al., 2020). For lesion-based

mass detection, the sensitivity of the 3D-Mask R-

CNN based CAD (segmentation) was 90% with 0.8

FPs/lesion, whereas the sensitivity of the Faster R-

CNN based CAD was 90% at 2.37 FPs/lesion. (Buda

et al., 2020) developed a single-phase deep learning

detection model for masses and architectural

distortions and achieved a sensitivity of 65% at 2

FPs/breast. (Li et al., 2021) propose a very interesting

work on Faster R-CNN that uses mammary gland

distribution as a prior information to detect

architectural distortions in DBT and achieved a

sensitivity of 80% at 1.85 FPs/volume for all

architectural distortions types.

A fair and direct comparison between our results

and these published data is not possible because they

analyze completely different lesions, those are

already optimized studies and of different

characteristics (for example, some use ROIs and not

the whole image to locate the lesions). Furthermore,

although architectural distortions are quite difficult to

locate, masses are more reasonable. Although masses

have densities similar to the rest of the breast tissue

and are often camouflaged, they are larger than

microcalcifications, facilitating training and learning.

It is possible to use images with less resolution and

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering

train more complex networks faster. Thus, we cannot

make a comparison between our results and those

already published, but we can conclude that, despite

our high FP values in this preliminary study, there is

potential to improve and achieve results similar to

those of the masses.

In conclusion, taking into account the preliminary

results presented, we conclude that detection and

location of MCs in DBT can be automatically

achieved using Faster R-CNN and visualization of

these results can benefit from another approach such

as 3D VR.

ACKNOWLEDGEMENTS

This work was supported by Universidade de Lisboa

(PhD grant) and Fundação para a Ciência e

Tecnologia – Portugal (Grant No.

SFRH/BD/135733/2018 and FCT-IBEB Strategic

Project UIDB/00645/2020).

REFERENCES

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

Samuelson, F. W., Myers, K. J. (2018). Evaluation of

Digital Breast Tomosynthesis as Replacement of Full-

Field Digital Mammography Using an In Silico

Imaging Trial. JAMA Network Open, 1(7), e185474-

e185474. doi: 10.1001/jamanetworkopen.2018.5474

Bernardi, D., Macaskill, P., Pellegrini, M., Valentini, M.,

Fantò, C., Ostillio, L., Houssami, N. (2016). Breast

cancer screening with tomosynthesis (3D

mammography) with acquired or synthetic 2D

mammography compared with 2D mammography

alone (STORM-2): a population-based prospective

study. The Lancet Oncology, 17(8), 1105-1113. doi:

https://doi.org/10.1016/S1470-2045(16)30101-2

Buda, M., Saha, A., Walsh, R., Ghate, S., Li, N., Święcicki,

A., Mazurowski, M. A. (2020). Detection of masses and

architectural distortions in digital breast tomosynthesis:

a publicly available dataset of 5,060 patients and a deep

learning model. arXiv:2011.07995. Retrieved from

https://ui.adsabs.harvard.edu/abs/2020arXiv20110799

Bunch, P., Hamilton, J., Sanderson, G., & Simmons, A.

(1977). A Free Response Approach To The

Measurement And Characterization Of Radiographic

Observer Performance (Vol. 0127): SPIE.

Caumo, F., Zorzi, M., Brunelli, S., Romanucci, G., Rella,

R., Cugola, L., Houssami, N. (2018). Digital Breast

Tomosynthesis with Synthesized Two-Dimensional

Images versus Full-Field Digital Mammography for

Population Screening: Outcomes from the Verona

Screening Program. Radiology, 287(1), 37-46. doi:

10.1148/radiol.2017170745

Fan, M., Li, Y., Zheng, S., Peng, W., Tang, W., & Li, L.

(2019). Computer-aided detection of mass in digital

breast tomosynthesis using a faster region-based

convolutional neural network. Methods, 166, 103-111.

doi: https://doi.org/10.1016/j.ymeth.2019.02.010

Fan, M., Zheng, H., Zheng, S., You, C., Gu, Y., Gao, X., Li,

L. (2020). Mass Detection and Segmentation in Digital

Breast Tomosynthesis Using 3D-Mask Region-Based

Convolutional Neural Network: A Comparative

Analysis. [Original Research]. Frontiers in Molecular

Biosciences, 7(340). doi: 10.3389/fmolb.2020.599333

Fenton, J. J., Taplin, S. H., Carney, P. A., Abraham, L.,

Sickles, E. A., D'Orsi, C., Elmore, J. G. (2007).

Influence of Computer-Aided Detection on

Performance of Screening Mammography. New

England Journal of Medicine, 356(14), 1399-1409. doi:

10.1056/NEJMoa066099

Food and Drug Administration (FDA) U.S. . (2013).

Premarket Approval application supplement for the

Selenia Dimensions 3D System Retrieved May, 2021

Fotin, S., Yin, Y., Haldankar, H., Hoffmeister, J., &

Periaswamy, S. (2016). Detection of soft tissue

densities from digital breast tomosynthesis:

comparison of conventional and deep learning

approaches (Vol. 9785): SPIE.

Freer, P. E., Riegert, J., Eisenmenger, L., Ose, D., Winkler,

N., Stein, M. A., Hess, R. (2017). Clinical

implementation of synthesized mammography with

digital breast tomosynthesis in a routine clinical

practice. Breast Cancer Research and Treatment,

166(2), 501-509. doi: 10.1007/s10549-017-4431-1

Gilbert, F. J., Tucker, L., Gillan, M. G. C., Willsher, P.,

Cooke, J., Duncan, K. A., Duffy, S. W. (2015). Accuracy

of Digital Breast Tomosynthesis for Depicting Breast

Cancer Subgroups in a UK Retrospective Reading Study

(TOMMY Trial). Radiology, 277(3), 697-706. doi:

10.1148/radiol.20 15142566

Girshick, R. (2015). Fast R-CNN. Paper presented at the

Proceedings of the IEEE international conference on

computer vision.

Girshick, R., Donahue, J., Darrell, T., & Malik, J. (2014).

Rich feature hierarchies for accurate object detection

and semantic segmentation. Paper presented at the

Proceedings of the IEEE conference on computer vision

and pattern recognition.

Good, W. F., Abrams, G. S., Catullo, V. J., Chough, D. M.,

Ganott, M. A., Hakim, C. M., & Gur, D. (2008). Digital

breast tomosynthesis: a pilot observer study. AJR Am J

Roentgenol, 190(4), 865-869. doi: 10.2214/ajr.07.2841

Gur, D., Abrams, G. S., Chough, D. M., Ganott, M. A.,

Hakim, C. M., Perrin, R. L., Bandos, A. I. (2009).

Digital breast tomosynthesis: observer performance

study. AJR Am J Roentgenol, 193(2), 586-591. doi:

10.2214/ajr.08.2031

Hofvind, S., Hovda, T., Holen, Å. S., Lee, C. I., Albertsen,

J., Bjørndal, H., Skaane, P. (2018). Digital Breast

Tomosynthesis and Synthetic 2D Mammography

versus Digital Mammography: Evaluation in a

Population-based Screening Program. Radiology,

287(3), 787-794. doi: 10.1148/radiol.2018171361

BIOIMAGING 2022 - 9th International Conference on Bioimaging

. ImageNet. (2021) Retrieved October, 2021, from

http://www.image-net.org

Katzen, J., & Dodelzon, K. (2018). A review of computer

aided detection in mammography. Clinical Imaging, 52,

305-309. doi: https://doi.org/10.1016/j.clinimag.20

18.08.014

Lai, X., Yang, W., & Li, R. (2020). DBT Masses Automatic

Segmentation Using U-Net Neural Networks. Comput

Math Methods Med, 2020, 7156165. doi:

10.1155/2020/7156165

Lång, K., Andersson, I., Rosso, A., Tingberg, A., Timberg,

P., & Zackrisson, S. (2016). Performance of one-view

breast tomosynthesis as a stand-alone breast cancer

screening modality: results from the Malmö Breast

Tomosynthesis Screening Trial, a population-based

study. Eur Radiol, 26(1), 184-190. doi:

10.1007/s00330-015-3803-3

Lehman, C. D., Wellman, R. D., Buist, D. S. M.,

Kerlikowske, K., Tosteson, A. N. A., Miglioretti, D. L.,

& Consortium, f. t. B. C. S. (2015). Diagnostic

Accuracy of Digital Screening Mammography With

and Without Computer-Aided Detection. JAMA

Internal Medicine, 175(11), 1828-1837. doi:

10.1001/jamainternmed.2015.5231

Li, Y., He, Z., Lu, Y., Ma, X., Guo, Y., Xie, Z., Chen, H.

(2021). Deep learning of mammary gland distribution

for architectural distortion detection in digital breast

tomosynthesis. Physics in Medicine & Biology, 66(3),

035028. doi: 10.1088/1361-6560/ab98d0

MathWorks. (2021). MATLAB trainFasterRCNN

ObjectDetector Retrieved October, 2021, from

https://www.mathworks.com/help/vision/

ref/trainfasterrcnnobjectdetector.html

Mota, A. M., Clarkson, M., Orvalho, L., Almeida, P., &

Matela, N. (2020). Calculation of transfer functions for

volume rendering of breast tomosynthesis imaging.

Paper presented at the 15th International Workshop on

Breast Imaging (IWBI2020), Leuven, Belgium.

Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-

CNN: Towards real-time object detection with region

proposal networks. Advances in neural information

processing systems, 28, 91-99.

Samala, R. K., Chan, H.-P., Hadjiiski, L., Helvie, M. A.,

Wei, J., & Cha, K. (2016). Mass detection in digital

breast tomosynthesis: Deep convolutional neural

network with transfer learning from mammography.

Med Phys, 43(12), 6654-6654. doi: 10.1118/1.4967345

Schroeder, W., Martin, K., & Lorensen, B. (2006). The

Visualization Toolkit: An Object-oriented Approach to

3D Graphics (4rd ed.). USA: Kitware.

Sechopoulos, I., Teuwen, J., & Mann, R. (2020). Artificial

intelligence for breast cancer detection in

mammography and digital breast tomosynthesis: State

of the art. Seminars in Cancer Biology. doi:

https://doi.org/10.1016/j.semcancer.2020.06.002

Suetens, P. (2009). Medical image analysis Fundamentals

of Medical Imaging (2nd ed., pp. 159-189). New York:

Cambridge University Press.

Sung, H., Ferlay, J., Siegel, R. L., Laversanne, M.,

Soerjomataram, I., Jemal, A., & Bray, F. (2021). Global

Cancer Statistics 2020: GLOBOCAN Estimates of

Incidence and Mortality Worldwide for 36 Cancers in

185 Countries. CA Cancer J Clin, 71(3), 209-249. doi:

https://doi.org/10.3322/caac.21660

Tabár, L., Dean, P. B., Chen, T. H.-H., Yen, A. M.-F., Chen,

S. L.-S., Fann, J. C.-Y., Duffy, S. W. (2019). The

incidence of fatal breast cancer measures the increased

effectiveness of therapy in women participating in

mammography screening. Cancer, 125(4), 515-523.

doi: https://doi.org/10.1002/cncr.31840

Venson, J. E., Albiero Berni, J. C., Edmilson da Silva Maia,

C., Marques da Silva, A. M., Cordeiro d'Ornellas, M.,

& Maciel, A. (2017). A Case-Based Study with

Radiologists Performing Diagnosis Tasks in Virtual

Reality. Stud Health Technol Inform., 245, 244-248.

VICTRE. (2018). The VICTRE Trial: Open-Source, In-

Silico Clinical Trial For Evaluating Digital Breast

Tomosynthesis, 2021, from https://wiki.cancerimaging

archive.net/display/Public/The+VICTRE+Trial%3A+

Open-Source%2C+In-Silico+Clinical+Trial+For+Eval

uating+Digital+Breast+Tomosynthesis

VTK. (2020). Visualization Toolkit - VTK Retrieved

February, 2020, from http://www.vtk.org/

Zackrisson, S., Lång, K., Rosso, A., Johnson, K., Dustler,

M., Förnvik, D., Andersson, I. (2018). One-view breast

tomosynthesis versus two-view mammography in the

Malmö Breast Tomosynthesis Screening Trial

(MBTST): a prospective, population-based, diagnostic

accuracy study. The Lancet Oncology, 19(11), 1493-

1503. doi: 10.1016/s1470-2045(18)30521-7

Detection of Microcalciﬁcations in Digital Breast Tomosynthesis using Faster R-CNN and 3D Volume Rendering