Abstract:

Breast cancer is one of the most famous mortal diseases facing the world. People in
both developed and developing countries are equally struggling with it, when the
cancer spreads in the other part of body, it will be unstoppable and in that moment
there is a brief time before cancer kills the patient.

The only way to eliminate this malignant disease is by detecting and treating it earlier
before symptoms accrue. Doctors advise people especially women over 40 years to
pass a mammography test periodically every two years in goal for detecting breast
cancer earlier and give the more chance for the treatment. Unfortunately, there is a coy
reluctance among people concerning the passage of this test, there are two principal
causes including the people’s fear from exposure to X-ray radiation especially in the
chest area and fees of testing. For minimize fees and make sure that mammography is
in the lowest radiation does, I propose in this project to make combination between
computer vision using deep convolutional neural networks (DCNNs) and dark field
mammography imaging.,

The (DCNNs) method improves the ability of the machine to detect, recognize and
classify cancer automatically from mammograms so we do not need expert doctors to
making decisions. By results we minimize fees, and by using dark field imaging
technique we minimize the radiation dose. Furthermore, the mammography resolution
is high, in the other hand we optimize fees cause dark field mammography technique
doesn’t need expensive equipments.

Key words:
Breast cancer, dark field imaging, low dose mammography, convolutional neural
networks, auto detection, x rays, non destructive testing.

Title:
Automated detection of breast cancer with dark field
mammography using deep convolutional neural
networks

Bouzid Yassine

[email protected]

1
Table of content:

CONTENT PAGE

Abstract ……………………………………………………...……………………………………………..….....1

Key words ………………………………………………..………………………………………………..….....1

Introduction………………..……………………………………………………………………………..….....3

Objectives ………………………………………………………………………………………………....….....5

Litirature review ……………………………………………………………………………………..….....5

Methodology ……………...…………………………………………………………………………..…….....6

Phase 1: Breast cancer………………………………………………………...…………………..……......7

Phase 2: Dark field mammography technique……….…………………………………..……......7

Phase 03: Deep conventional neural network……………..……………………………..……......8

Conclusion ………………………………………………………………………………………..……..............9

Provisory time table for this project: …………………………………………………..…….........9

References………………………………………………………………………………………..……................9

2
Introduction:
Breast cancer is the most common cancer in women in both developed and developing
countries and the second most common cancer in the world. Low- and middle-income
countries (LMICs) represent most of the countries with the highest mortality rates,
ranging from 40% to 60%. Five-year survival rates for breast cancer are much worse
for LMICs countries such as Brazil, India, and Algeria in comparison with the United
States and Sweden.

More women are diagnosed with breast cancer than any other cancer, excluding skin
cancer. This year, an estimated 266,120 women in the United States will be diagnosed
with invasive breast cancer, and 63,960 women will be diagnosed with in situ breast
cancer. An estimated 2,550 men in the United States will be diagnosed with breast
cancer.[13]

It is estimated that 41,400 deaths (40,920 women and 480 men) from breast cancer will
occur this year (2018)

Since 1989, the number of women who have died of breast cancer has steadily
decreased thanks to treatment improvements and the early detection of this cancer.

Mammography plays an important role to detect abnormalities in the breast. It is

currently the most effective imaging modality for breast cancer screening.

A mammogram is an x-ray image of the breast that can reveal abnormalities like
benign or malignant while Screening Mammogram is performed to attempt to detect
breast cancer before symptoms occur. The goal of screening mammography programs
is the early detection of the cancer which decreases mortality from this cancer.

Unfortunately, the rate of early detection of breast cancer is estimate low especially in
developing countries; the fear of absorption of x rays radiation and the high price of
mammography examination is the most responsible causes for provoking people to
refrain from doing mammography before appearing the symptoms, for solving this
problem I propose in this project to combine between the automatic computer vision
for minimize fees of the examination with a low dose mammography technique.

En general, conventional radiography is based on the attenuation of x rays for

generating contrast which is natively low for the soft-tissue components of the female
breast.

Dark-field images are produced by small-angle scattering of x-ray photons. This

scattering signal has good sensitivity to structural variations and density fluctuation on
a length of several nanometers; it is intrinsically capable of detecting subtle differences
in the electron density of a material and of measuring the effective integrated local

3
small-angle scattering power generated by the microscopic density fluctuations in the
specimen

Dark-field mammography is superior to standard breast imaging modalities providing

a better resolution of small calcified tumor nodules, demarcation of tumor boundaries
by the illumination of its edge which provides a quantitative attenuation representation
of the soft-tissue sample using a simple conventional rotating anode x ray tube. [4]

It exhibits negligible requirements in terms of temporal or spatial coherence, it

provides high sensitivity, it is also robust against mechanical and thermal instabilities
and enables low-dose implementations of x ray dark field imaging in planar and three-
dimensional imaging.[2] this advantages make this technique the most suitable for
realizing computer vision of micro-calcification and cancers from the mammograms,
the edge illumination technique and the clear boundaries in this technique save the
information of the micro-calcification which may lost on the electronic segmentation or
the image thresholding before the detection of the cancer.

Since 2012, deep learning has transformed computer vision and has been implemented
in different domain as in scores of consumer-facing products and clinical applications.

on April 11, 2018, an important step was taken towards this future: the U.S. Food and
Drug Administration announced the approval of the first computer vision algorithm
that can be utilized for medical diagnosis without the input of a human clinician [5]
this step made researchers more excited to go forward in deep learning techniques for
fulfill the fully automated computer vision.

Convolutional neural network (CNN) is a specific type of artificial neural network that
uses a machine learning unit algorithm for supervised learning, to analyze data called
perceptrons. Deep Convolutional Neural Networks (DCNNs) is about finding weights
using algorithms that make the (CNN) exhibit desired behavior, perceptrons showed
promising performance as well speed in different domains like: speech recognition,
text recognition, lips reading, Computer-Aided Diagnosis(CAD), face recognition,
drug discovery …etc (DCNNs) have pushed the performance of computer vision
systems to soaring heights on a broad array of high-level problems, including image
classification and object detection, where DCNNs trained in an end-to-end manner
have delivered strikingly better results than systems relying on hand-crafted
features.[7]

The efficiency of the (DCNNs) in classification and recognition allow me to propose

(DCNNS) approach for detect and localize the micro-calcification in the dark field
mammograms for achieving the early detection of breast cancer and fulfill the
classification either the cancer is benign or malignant using specific algorithms

4
Objectives:
The objective of this work is to combine between dark field imaging mammography
and deep convolutional neural network to create an automated machine vision by
supervised learning

Literature review:
The origin of mammography can be traced to the discovery of x-rays by Wilhelm
Röntgen in 1895[14], He was awarded the first Nobel Prize in Physics in 1901 [12] In
the late 1950s Robert Egan at the University of Texas M.D. Anderson Cancer
Center combined a technique of low kVp with high mA and single emulsion films to
devise a method of screening mammography for the first time. He published these
results in 1959 in a paper, and subsequently in a 1964 book called Mammography

In general, cancers start when cells begin to grow out of control. These cells usually
form a foreign tissue in the body called tumor that can often be seen on an x-ray or felt
as a lump. The tumor is malignant (cancer) if the cells can grow into (invade)
surrounding tissues or spread (metastasize) to distant areas of the body. Breast cancer
is one type of cancers it can start from different parts of the breast. Most breast cancers
begin in the ducts or in the glands [10].
The development of the mammography for decades produce many techniques and
methods for enhancing the breast tissue images like the phase contrast imaging, the
principle of this method in general was developed by Frits Zernike during his work
with diffraction gratings and visible light December 1942 [3]. The implementation of
the phase-contrast method in X-ray physics was presented until 1965 by Ulrich Bonse
and Michael Hart, due to the slow progress in improving the quality of X-ray beams
and the non-availability of X-ray optics (lenses), after that a new approach appear
called grating-based imaging, which makes use of the Talbot effect, discovered by
Henry Fox Talbot in 1836 [11].
The first X-ray grating interferometers using synchrotron sources were developed in
2002 by Christian David and colleagues from the Paul Scherrer Institute (PSI) in
Villingen, Switzerland[1]
In 2006, a great advancement was the transfer of the grating-based technique to
conventional laboratory X-ray tubes by Franz Pfeiffer and co-workers [9] which fairly
enlarged the technique's potential for clinical use.
In 2008 the group of Franz Pfeiffer also accomplished to extract a supplementary
signal from their experiments, the so-called "dark-field signal" was caused by
scattering due to the porous microstructure of the sample and provided
"complementary and otherwise inaccessible structural information about the specimen
at the micrometer and submicrometer length scale"[8]. Before 2012 researchers use the
synchrotron radiation with the crystal interferometer for this technique but Han Wen
and co-workers in 2012 took a step forward by replacing the crystals with nanometric
phase gratings which allow the use conventional X-ray tubes [15]

5
Most recently, a significant advance in grating-based imaging occurred due to the
discovery of a phase moiré effect which allow imaging at low radiation doses to
achieve high sensitivity [6], by reviewing this optical x ray techniques which their
main goal is the enhancement of the brute mammogram digital image we take a look at
the numerical techniques for the auto detection, recognition and classification of breast
cancer from mammograms.

In general, we saw that the enhancement of the brute medical images which was been
taken is indispensable for the good diagnostic and treatment, the research was be
incredible increasing, machine learning techniques (ML) have played an important role
in medical field like medical image processing, computer-aided diagnosis (CAD),
image interpretation, image fusion, image registration, image segmentation…etc.

These techniques enhance the abilities of doctors and researchers to understand that
how to analyze the generic variations which will lead to disease. These techniques
composed of conventional algorithms without learning like Support Vector Machine
(SVM), Neural Network (NN)…etc, and recently, an ML area called “deep learning”
emerged in the computer vision field and became very popular in many fields, since
2012, when a deep-learning approach based on a convolutional neural network (CNN)
won an overwhelming victory in the best-known worldwide computer vision
competition, ImageNet Classification. There are many other deep learning algorithms
appear after that such as Recurrent neural Network (RNN), Long Short term Memory
(LSTM), Extreme Learning Model (ELM), Generative Adversarial Networks (GANs)
etc [7].

Automated detection of diseases based on conventional methods in medical imaging

has been shown significant accuracies in computer vision which give us the
opportunity to use one of these techniques in goal for achieving the aims of this
research project.

Methodology:
I think keeping always up to date and separate theory and the experimental work
will give more accurate work, straight and organized which helps to attain the
main goal of each section in this the project.

For achieving the objectives of this project, I will propose to divide work on three
phases the first concerning the breast cancer and mammography, the second regarded
the x rays dark field imaging technique, the last one will be specialized for the deep
conventional neural networks (DCNNs) each phase contain two sections, theoretical
and experimental.

6
Phase 1: Breast cancer
Section 1:

For the first phase it is very important to collect maximum information about this kind
of tumors and mammography in general, so I need to read the maximum of scientific
books and documents about the newest research concerning breast cancer especially
focusing on its types, proprieties, causes and what it is the first ingredients
constitutions. For this, I will use all the available recourses such as hospitals, clinics,
interviews with doctors and patients... etc. The most important source is recent
researches about this cancer, being a member at a university or at a research institute
will give me the ability to access in most of research libraries around the world.

Section 2:

After collecting most important information about this cancer, I will start with
mathematic modeling by trying to create the mathematic model corresponding to the
breast tissue and cancer proprieties and its reactions with the x rays model, this step is
very important because if I get the nearest model to the real reaction, I will gain the
opportunity to simulate and try many experiments on the mathematic model without
funding, this will minimize the time and fees of research.

I will use Matlab for this section because I had used it very much before in
undergraduate studies so I am very familiar with this tool, the experiment test to prove
the mathematic model efficiency will be done by comparing the simulation results with
the real image given by real breast tissue and cancer, according to the results of
comparison, I will enhance and update the model until it gives the desired results, this
will be my first paper.

Phase 2: Dark field mammography technique

Section 01:

Next, I will start again theory work, in this section will be dedicated to grating
imaging, interferometer, phase contrast imaging and dark field imaging technique, I
will study the proprieties of the dark field imaging technique and its interaction with
the breast tissue especially with early calcifications, the dark background formed by the
dark field imaging technique gives a very good contrast and makes boundaries
optically very clear before any numerical processing on the image (figure 1)

Section 02:

Dark field imaging technique is very useful to decrease the error rate of thresholding
without losing information; this propriety is indispensible for detecting tiny
calcification in soft tissue as the one of the breast.

7
 My experience for years in industrial nondestructive x ray quality control helps me to
understand more about this technique in experimental work. Wisdom supervising
requests and the laboratory’s equipments will give me the opportunity to make very
well research about this technique and give me the chance for practicing experiments
using this technique; in this section, I will publish the results of experiments that I will
do.

Figure 1: (a) experimental absorption-contrast mammography (b) dark-field mammography. [4]

Phase 03: Deep conventional neural network

Section 01: After that I will go to theoretical research for deep machine learning techniques,
by analyzing data from the two previous phases we have two levels of classification. First, if
the image contains a cancer or not then if this cancer is benign or malignant, I use to take
advantage of my experience in machine learning particularly Convolutional Neural Network
(CNN), because I have already used to solve reprise problem of the power system in my master
final project, so I have initial basic ideas about this method

Section 02: For the experimental work, I will improve techniques of image processing
and segmentation including extraction algorithm, then, I will build the model of
(DCNNs) and train it, using algorithms to recognize the shape, asymmetry and density
of cancer from mammograms, until the model will get the value of weights desired and
exhibit the recognition of shape, asymmetry and density of cancer, then I will start
another training of (DCNNs) model to get weights able to make difference between
benign and malignant cancer, which needs a robust algorithm with special model, I
will work for it very carefully in the last section. Finally I will make a comparison
between images taken by the conventional transmission and the others by dark field
scattering imaging according to the error rate of the (DCNNs) model.

In the end I will prove that the dark field imaging technique is very useful for deep
neural networks model and both together will solve the problem of this project.

8
Dark field imaging technique and deep convotional neural networks (DCNNs)
technique will help us to make sheep mammograms with low dose, high image quality
and negligible error, according to the results of this experiments and analysis I will
write my thesis in the rest of time

Conclusion:
The proprieties of dark field imaging mammography allow us to do a safe
mammography with high imaging quality and without huge requirements like
synchrotron. In the same time, with low radiation dose which makes people feel safe
with mammography exam, then the deep convolutional neural network (DCNNs) helps
us to create a good computer vision which helps people to get the mammography
results with no necessity of expert doctor’s inspection which will absolutely minimize
fees of each exam

We going one step forward on the way of fully automatic clinical mammography and doing a
good advancement against this malignant disease. In the end, the most important goal will be
attained by achieving aims of this project that we will save many lives.

Provisory time table for this project:

Work desired Period
Phase 01: Breast cancer Section 01 Three months.
Section 02 Six months.
Vacation One month
Phase 2: Dark field mammography Section 01 Two months.
Section 02 Six months.
Vacation One month
Phase 3: Deep convolutional neural network Section 01 Four months.
Section 02 Six months.
Writing thesis and supplement work six months

References:

[1] David, C., Nöhammer, B., Solak, H. H., & Ziegler, E. (2002). Differential x-ray
phase contrast imaging using a shearing interferometer. Applied Physics Letters,
81(17), 3287–3289. doi:10.1063/1.1516611

[2] Endrizzi, M., Vittoria, F. A., Rigon, L., Dreossi, D., Iacoviello, F., Shearing, P. R.,
& Olivo, A. (2017). X-ray Phase-Contrast Radiography and Tomography with a
Multiaperture Analyzer. Physical Review Letters, 118(24).
[3] F. Zernike, Groningen.(1942). Phase Contrast, A New Method for the Microscopic
Observation of Transparent Object part II.
[4] Grandl, S., Scherer, K., Sztrókay-Gaul, A., Birnbacher, L., Willer, K., Chabior, M.,
… Hellerhoff, K. (2015). Improved visualization of breast cancer features in multifocal

9
carcinoma using phase-contrast and dark-field mammography: an ex vivo study.
European Radiology, 25(12), 3659–3668. doi:10.1007/s00330-015-3773-5
[5] Keane, P. A., & Topol, E. J. (2018). With an eye to AI and autonomous diagnosis.
Npj Digital Medicine, 1(1). doi:10.1038/s41746-018-0048-y

[6] Miao, H., Gomella, A. A., Harmon, K. J., Bennett, E. E., Chedid, N., Znati, S., …
Wen, H. (2015). Enhancing Tabletop X-Ray Phase Contrast Imaging with Nano-
Fabrication. Scientific Reports, 5(1). doi:10.1038/srep13581

[7] Muhammad Imran Razzak, Saeeda Naz, Ahmad Zaib. Deep Learning for Medical
Image Processing: Overview, Challenges and Future.
[8] Pfeiffer, F., Bech, M., Bunk, O., Kraft, P., Eikenberry, E. F., Brönnimann, C.,
David, C. (2008). Hard-X-ray dark-field imaging using a grating interferometer. Nature
Materials, 7(2), 134–137. doi:10.1038/nmat2096

[9] Pfeiffer, F., Weitkamp, T., Bunk, O., & David, C. (2006). Phase retrieval and
differential phase-contrast imaging with low-brilliance X-ray sources. Nature Physics,
2(4), 258–261. doi:10.1038/nphys265

[10] Smith, R. A., Andrews, K. S., Brooks, D., Fedewa, S. A., Manassaram-Baptiste,
D., Saslow, D., … Wender, R. C. (2017). Cancer screening in the United States, 2017:
A review of current American Cancer Society guidelines and current issues in cancer
screening. CA: A Cancer Journal for Clinicians, 67(2), 100–121.
doi:10.3322/caac.21392

[11] Talbot, H. F. (1836). LXXVI. Facts relating to optical science. No. IV. The
London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 9(56),
401–407. doi:10.1080/14786443608649032

[12]The Nobel Prize in Physics.(1901).Nobelprize.org

[13] U.S. Breast Cancer Statistics (January 9, 2018).breastcancer.org

https://www.breastcancer.org/symptoms/understand_bc/statistics

[14] W. C. Röntgen.(Feb. 14, 1896). On a New Kind of Rays

American Association for the Advancement of Science, Vol. 3, No. 59, pp. 227-231

[15] Wen, H., Gomella, A. A., Patel, A., Lynch, S. K., Morgan, N. Y., Anderson, S. A.,
… Wolfe, D. E. (2013). Subnanoradian X-ray phase-contrast imaging using a far-field
interferometer of nanometric phase gratings. Nature Communications, 4(1).
doi:10.1038/ncomms3659

[16]Yang, X., De Andrade, V., Scullin, W., Dyer, E. L., Kasthuri, N., De Carlo, F., &
Gürsoy, D. (2018). Low-dose x-ray tomography through a deep convolutional neural
network. Scientific Reports, 8(1). doi:10.1038/s41598-018-19426-7
doi:10.1103/physrevlett.118.243902

10