INTRODUCTION.

Breast cancer is one of the most common cancers diagnosed in women around the
world and it is a main cause of fatality among women. In low-income and middle-
income countries the mortality rates are relatively high compared to developed coun-
tries. According to the World Health Organization’s International Agency for Research
on Cancer (IARC) 2013 report, more than 1.7 million women in 2012 were diagnosed
with breast cancer world- wide. This is considered around 11.9% of all cancers
diagnosed in the same year with 522,000 death cases reported. It is also ex- pected that
by 2025 there will be 19.3 million new cancer cases. Moreover, in developing countries
like Egypt, the dense pop- ulation and the patients’ ignorance to the disease symptoms
and seeking medical consultation either when it’s too late or extremely critical leads to
higher mortality. Also, shortage of medical special- ists and experts in rural areas adds
up to the problem of early and accurate diagnosis of breast cancer causing higher
mortality rate.

Machine learning techniques

Reference Imaging Machine Scope Evaluation Advantages

modality learning results Image set Disadvatages
technique results
[1] DM SVM breast mass AUC = 600 ➨SVM’s are ➨Choosing a
classification 0.805 ± malignant very good “good” kernel
0.012 when we have function is not
no idea on the easy.
data. ➨Long training
➨Works well time for large
with even datasets.
unstructured
and semi
structured data
like text,
Images and
trees
[2] DM KNN classifying ROI AUC = 0.9713 MIAS ➨No Training ➨Does not
as normal or 109 cases Period. work well with
abnormal ➨Since the KNN large dataset.
algorithm requires ➨Does not
no training before
work well with
making predictions,
new data can be high
added seamlessly dimensions.
which will not
impact the
accuracy of the
algorithm.

[3] DM ANN multiple SVM: AUC = DDSM ➨Storing ➨Hardware

SVM classifier system 0.932 303 images information on the dependence:
entire network : Artificial neural
for masses ANN: AUC = Information such as networks require
classification 0.925 in traditional processors with
programming is parallel processing
stored on the power, in
entire network, not accordance with
on a database. The their structure. For
disappearance of a this reason, the
few pieces of realization of the
information in one equipment is
place does not dependent.
prevent the
network from
functioning.
[4] IRT ACO classifying Acc = Private ➨Have ➨Have
breast into 79.52% 146 images granted difficult
benign and 29 malignant coverage. theoretical
malignant cases 117 benign ➨Can adapt to analisys.
changes. ➨Have
uncertain time
to coverage.
microcalcifications
[5] DM Fuzzy-C- Acc = MIAS ➨ Gives best ➨ Apriori
cluster enhancement
Means result for specification of
method 95%, Sn =
(FCM) 93% overlapped the number of
MIAS: data set and clusters.
Acc = comparatively ➨With lower
better then k- value of β we
94%, Sn =
82% means get the better
algorithm. result but at
➨ Unlike k- the expense of
means where more number
data point must of iteration.
exclusively
belong to one
cluster center
here data point
is assigned
membership
to each cluster
center as a
result of which
data point may
belong to more
then one
cluster center.
[6] US LDA distinguishing AUC = 0.85 Private 90 ➨It uses ➨Does not
positive and patients information work well if the
negative lymph From both the design is not
nodes. features to balanced.
create a new LDA is sensitive
axis which in to over fit and
turn minimizes validation od
the variance LDA model are
and maximizes least
the class problematic.
between two
variables.
[7] US LR Classification oF Sn = 95%, Sp Private ➨It can also be ➨The
BI-RADS = 73%, 69 masses extended from interpretation
AUC = 0.95 21 malignant binary is more
48 benign classification to difficult
multi-class because the
classification. interpretation
On the good side, of the weights
the logistic is
regression model multiplicative
is not only a
and not
classification
additive.
model, but also
gives you
probabilities.
[8] Histopathology DL Breast cancer Acc = 0.86 ± Private 58 H&E- ➨Features are ➨It requires very
automatically large amount of data
detection 0.03, Sn = 1, stained deduced and in order to perform
Sp = 0.72 ± histopathology optimally tuned for better than other
desired outcome. techniques.
0.10 images of Features are not ➨It is extremely
breast tissue required to be expensive to train
extracted ahead of due to complex data
time. This avoids time models. Moreover
consuming machine deep learning
learning techniques. requires expensive
➨The deep learning GPUs and hundreds
➨architecture is of machines. This
flexible to be adapted increases cost to the
to new problems in users.
the future. ➨There is no
standard theory to
guide you in selecting
right deep learning
tools as it requires
knowledge of
topology, training
method and other
parameters. As a
result it is difficult to
be adopted by less
skilled people.
[9] US SVM Clasifing solid ACC = private ➨Better accuracy ➨Computationally
75.5, Sn as compare to expencive.
breast cancer = 110 images
other classifier. ➨SVM was
78.9%, designed to solve
Sp = 73.6%, ➨Overfitting
the problem of
AUC = 0.82 problem is not as binary class.It
much as others solves by breaking
methods. it into pair of two
classes such as
one-against-one
and one-against all

[10] DM LR Classify breast AUC=0.7838 DDSM ➨It can also be ➨The

Private (1006 cases
cancer extended from interpretation
(646 benign and
 360 malignant)) binary is more
classification to difficult
multi-class because the
classification. interpretation
On the good side, of the weights
the logistic is
regression model multiplicative
is not only a
and not
classification
additive.
model, but also
gives you
probabilities.
[11] Microscopic RF Diagnosis of Acc = 90%, Sn Private 228 ➨There is no ➨They’re not
imges SVM breast cancer = 94.59%, Sp images need for easily
= 96.72% feature interpretable.
normalization. ➨They’re not
➨Individual a state-of-the-
decision trees art algorithm.
can be trained
in parallel.
➨Random
forests are
widely used.

[1] M. Tan, J. Pu, B. Zheng, Optimization of breast mass classification using sequential forward floating selection (SFFS) and a support vector machine
(SVM) model, Int. J. Comput. Assisted Radiol. Surg. 9 (6) (2014) 1005–1020.
[2] S.J.S. Gardezi, et al., Mammogram classification using dynamic time warping, Multimedia Tools Appl. 76 (282) (2017) 1–22.
[3] A. García-Manso, et al., Consistent performance measurement of a system to detect masses in mammograms based on blind feature extractio n,
Biomedm Engm Online 12 (1) (2013) 2.
[4] G. Schaefer, ACO classification of thermogram symmetry features for breast cancer diagnosis, Memetic Comput. 6 (3) (2014) 207–212.
[5] L. Vivona, et al., Fuzzy technique for microcalcifications clustering in digital mammograms, BMC Med. Imaging 14 (1) (2014) 23.
[6] K. Drukker, et al., Quantitative ultrasound image analysis of axillary lymph node status in breast cancer patients, Int. J. Comput. Assisted Radiol.
Surg. 8 (6) (2013) 895–903.
[7] W.K. Moon, et al., Quantitative ultrasound analysis for classification of BI-RADS category 3 breast masses, J. Digital Imaging 26 (6) (2013)
1091–1098.
[8] H. Rezaeilouyeh, A. Mollahosseini, M.H. Mahoor, Microscopic medical image classification framework via deep learning and shearlet transform,
J. Med. Imaging 3 (4) (2016) 044501.
[9] H.S. Tseng, et al., Speckle reduction imaging of breast ultrasound does not improve the diagnostic performance of morphology -based CAD system, J.
Clin. Ultrasound 40 (1) (2012) 1–6.
[10] H. Jing, Y. Yang, R.M. Nishikawa, Retrieval boosted computer-aided diagnosis of clustered microcalcifications for breast cancer, Med. Phys. 39
(2) (2012) 676–685.
[11] M.M. Fernández-Carroblesa, et al., A CAD system for the acquisition and clas- sification of breast TMA in pathology, Stud. Health Technol. Inf. 210
(2015) 756–760.