Breast Cancer Classification using Machine

Learning

Abstract this research, six classification models were utilized

in the Wisconsin Breast Cancer (original) data sets,
Breast cancer is the most common cancer among
both before and after utilizing the Key Segment
women, affecting around 10% of them at some time
Analysis: Decision Tree, K-Neighbours, Logistic
in their lives. The percentage has continued to rise in
Regression, Random Forest, and Vector Support
recent years, with data suggesting that the survival
Machine (SVM). The findings achieved are above
rate is 88 percent after five years and 80 percent after
and above what can be measured using multiple
ten years following diagnosis. Breast cancer early
performance measures to compare algorithms in
detection has improved significantly since 1989, with
order to find the best cancer prediction, model.
mortality rates for the disease now at 39%.
Mammography, ultrasound, and biopsy will be done It's a popular method for detecting the tumor's edges
to assess their chances of being diagnosed with breast from various perspectives. The pictures created
cancer. The most typical expression of these provide a radiologist with information about the
processes is a biopsy, which involves the removal of tumour and how it spread across the area. This
sample cells for examination. During the breast finale procedure takes time, and the outcomes are highly
aspiration (FNA) operation, a sample of cells is dependent on the staff's expertise and experience;
collected and sent to a pathology lab for microscopic also, specialists are not available in every part of the
inspection. Numerical features such as radius, texture, world. As a result, the research community developed
perimeter, and the area may all be measured with a computer-assisted diagnosis (CAD) system for
incredibly tiny images. The results of the FNA are better tumour categorization, accurate findings, and
then paired with imaging data to predict the chance of quick execution without the need for radiologists or
a patient developing aggressive breast cancer. specialists. Machine learning algorithms (MLs) are
proposed as a replacement for human vision and
Keywords – Breast Cancer, ML, Classification,
expertise in analysing medical pictures and making
Prediction
high-accuracy judgements.
Introduction
There are six key phases to using machine learning
In this case, the automated system can be really methods (i.e., loading images, pre-processing,
useful. It will most likely speed up the procedure and segmentation, features extraction, features selection,
increase the doctor's prognosis accuracy. and classification). The first three processes are in
Furthermore, if it is approved, it would likely charge of getting rid of any pixels that aren't
eliminate the need for patients to undergo additional connected to the tumour. Converting photos to
procedures such as mammography, ultrasound, and statistical characteristics and decreasing their size by
MRI, all of which increase the risk of sickness. picking the most relevant ones is the job of features
Radiation and pain Early prediction is a crucial aspect extraction and selection. Finally, one of the machine
of the tracking process in general. Data mining and learning algorithms is used to categorise datasets and
segregation techniques can assist to minimise the provide a final result. Deep learning has recently been
amount of bad and bad judgments. As a result, used to categorise breast cancers, which is a novel
innovative approaches like data gathering in archives method. This technique bypasses the pre-processing
have grown popular among medical researchers. In
and feature extraction processes, and a large dataset from a knowledge set to get groups or clusters and
can improve the method's speed. describe the word set itself. Bracket, also appertained
to as supervised literacy in machine literacy, aims to
Machine learning
classify unknown situations supported learning being
Machine learning is a type of learning in which a patterns and orders from the word set and latterly
machine may learn on its own without being prognosticate unborn situations. The training set,
explicitly programmed. It's a subset of artificial which is employed to produce the classifying
intelligence that's extremely crucial. Machine structure, and thus the test set, which tends to assess
Learning (ML) has recently gained popularity among the classifier, are generally mentioned in bracket
researchers, companies, and government sectors in tasks Bracket may be a relatively complex
the rapidly evolving technological world. It's a optimisation problem. Numerous ML ways are
method of automating and developing the learning applied by experimenters in working on this bracket
process of systems/computers based on their problem. The most notorious algorithm that's used for
capabilities/understanding capacities without bone cancer bracket or vaticination is an artificial
requiring human intervention. The machine learning neural network, arbitrary timber, support vector
process begins by providing raw data into machine machine, etc. Scientists strive to seek out the simplest
learning models or learning tasks to train the algorithm to realise the foremost accurate bracket
machines. To construct ML models or ML tasks, result, still, data of variable quality also will impact
many machine learning algorithms or methodologies the bracket result. Further, the oddity of knowledge
employ input data and intended output. Machine will impact the number of algorithm operations also.
learning techniques will be discussed, including Nave However, there are more treatment options and a far
Bayes, K-nearest neighbor, logistic regression, better chance for survival, If the melanoma is planned
reinforcement algorithm, and support vector machine beforehand. Women whose melanoma is detected at
algorithm. We'll spend most of our time looking at an early stage have a 93% or advanced survival rate
these different algorithms and analyzing their within the first five times. Chancing cancer
outcomes. Researchers can increase their beforehand can also save your life. Breast cancer is
understanding of the cause and demonstrate currently affecting a large number of people. The
correlations, such as how a cancer patient responds to exact cause of this disease is dependent on human
pharmacological therapy, to make clinical factors and cannot be discovered easily. Furthermore,
predictions. Another alternative is to look into the the process of determining whether the cancer is
connection.Machine learning, which is a complicated benign or malignant necessitates a great lot of work
technique of looking at correlations, may be used to on the part of doctors and physicians. When
analyze complex cancer biology data. Computers can numerous tests are involved in the detection of breast
find these linkages in very vast and complicated data cancer, such as clump thickness, cell size uniformity,
sets, which is a benefit. cell form uniformity, and so on, the final result might
Background be difficult to get, even for physicians. Diagnostic
computer-based apps are widely utilized to diagnose
The cancer bracket divides melanoma into orders illnesses that claim many lives. In operations rooms,
depending on how they've spread or if they've spread robotics plays an extremely important role. Within
at all. Bracket algorithms prognosticate one or further the intense treatment rooms, the skilled systems are
separate variables, supported the contrary attributes also given.
within the dataset. data processing software is needed
to run the bracket algorithms. the end of bracket is to Using another aspect of Artificial Intelligence for
pick the simplest treatment. Bracket is vital because breast cancer detection, on the other hand, isn't
it allows scientists to spot, group, and duly name worthless. Breast cancer is the second most frequent
organisms via a invariant system. Bracket and disease among women, and it was the most common
clustering are two extensively used styles in the data cancer in the globe by the year 2002, according to
processing. Clustering styles aim to prize information reports. This disease is a very frequent kind of cancer
in girls, and it is the second leading cause of cancer 40 to 70. However, due to the location of the disease
mortality. In the United States, one out of every eight or the thickness of the chest tissue, a mammography
females faces the possibility of acquiring breast cannot detect all malignancies. A screening
cancer over their lifetime. The uncontrolled division mammography can identify around 25% of cancers in
of one cell within the breast leads to the onset of women aged 40 to 49, compared to 10% of
breast cancer, which manifests itself as a visible lump malignancies in women aged 50 and over. Other
known as a tumor. It is possible for the tumor to be algorithms do not employ picture datasets and instead
benign or cancerous. The right classification in generate predictions based on inaccurate or
determining whether a tumor is benign or malignant incomplete data, which frequently leads to incorrect
can save lives. As a result, the need for exact findings. Although the use of mammography image
classification within the clinic might be a source of datasets provides accurate predictions, it does not
great anxiety among experts and clinicians. Artificial identify all malignancies. Other datasets fail to
intelligence has been growing in prominence over the perform well because either the dataset is inaccurate
past twenty-five years since scientists realized the or they employ a small number of features in training
value of making bound selections to cure certain that aren't as relevant in selecting BC in the real
ailments. The use of machine learning and data world.
processing as diagnostic tools has become extremely
The proposed method creates a classification and
effective, and one of the most important illnesses in
predictive model for breast cancer that will provide
medicine where the classification task plays a critical
accurate grouping and prediction. This suggested
role is a breast cancer diagnosis.
method would use ten real-world value factors for
As a result, machine learning algorithms will assist each cancer cell nucleus to forecast the illness at an
clinicians incorrectly identifying breast cancer and early stage. The suggested model is implemented and
classifying it as benign or malignant tumors. There is developed using a combination of classification and
no doubt that the analysis of patient data and the ensemble learning methods. Attributes are defined as
selection of doctors and specialists are the most properties of the cell nuclei seen in a digitised picture
important factors in the identification process, but of a fine needle aspirate (FNA) of a breast lump. The
intelligent systems and artificial intelligence qualities of the cell nuclei detected in the picture are
techniques such as machine learning for classification described by the features in the dataset. There are 569
tasks also help doctors and specialists a great deal. In observations and 32 observations, which provide the
this research, we examine several categorization patient's identification, diagnosis, and 30 significant
learning algorithms in terms of their ability to real-valued attributes.
distinguish benign from malignant cancer in a breast
Related works
cancer dataset. We want to look at several machine
learning approaches; therefore, we'll employ a variety McDonald and his colleagues used data from 499
of algorithms and apply them to a breast cancer independent cells from the National Cancer Institute
dataset. to develop machine-based models that could predict
15 distinct forms of cancer.The models were then
The System
tested against a clinical database including seven
Many current methods develop models and make different chemotherapy medicines that were given to
predictions using the notion of a mammogram picture 23 ovarian cancer patients one at a time or in
collection. An x-ray image of the breast is called a combination. The model was able to predict with a
mammogram. It's also helpful if you have a lump or total accuracy of 91%.
other cancer-related symptom. Screening
Study
mammography is a form of mammogram that is used
to monitor people who have no signs or symptoms of "Our initial discovery of 90% accuracy in predicting
breast cancer. It contributed to a decrease in the drug responses in cervical cancer patients is very
number of breast cancer deaths among women aged promising and gives me hope," McDonald says.
"While further confirmation will need to be made
using large numbers of patients with multiple Methods and Tasks
cancers," he adds, "our initial discovery of 90%
Health organizations have undertaken efforts to
accuracy in predicting drug responses in cervical
minimize the cost of treating breast cancer and to
cancer patients is very promising and gives me hope."
improve the quality of treatment provided to these
Drug reactions are difficult to anticipate due to cancer
patients. Some of the diagnostic and laboratory
consequences. Even people with the same form of
procedures used to treat breast cancer are costly and
cancer may respond differently to the same treatment,
inconvenient for patients. Furthermore, certain
according to McDonald's."Part of the issue is that a
medications that were formerly thought to be safe
cancer cell is a highly integrated network of patient
have since been found to have a variety of negative
pathways, and tumors with comparable clinical
effects after long-term use. As a result, the general
characteristics might have considerable differences at
public has grown used to data mining techniques
the cellular level," McDonald stated. The objective of
being used to treat chronic diseases like breast cancer.
personalized cancer care is to precisely anticipate
My goal in this work is to use characteristics extracted
responses based on each patient's tissue's genetic
from numerous cell pictures to identify tumors as
profiles.
malignant (cancerous) or benign (non-cancerous). A
Approach digitized picture of a fine needle aspirate (FNA) of a
breast mass is used to compute features. They define
"In our approach, we use machine learning
the features of the image's cell nuclei. SVM is a
approaches to develop predictable algorithms in
supervised machine learning technique that may be
different cancer treatments based on correlations
used to solve issues like classification and regression.
between gene expression genes of cancer cells or
It transforms your data using a method known as the
patient tumors with predicted responses." The long-
kernel trick and then calculates an ideal boundary
term objective is to include tumor samples' genetic
between the available outputs based on these
profiles into algorithms so that prospective patient
alterations.
responses to various pharmacological therapies may
be predicted with high accuracy," McDonald 1. Traditional Models
added.Until this research, most machine learning
studies used a single algorithmic technique. Logistic Regression
McDonald and his colleagues, on the other hand, The likelihood of the default class (in this example,
choose to apply a few algorithm methods that Class 2) is predicted using logistic regression, which
combine numerous data analysis techniques. turns the probability into a binary value (0 or 1) for
Researchers discovered that using this term-based classification using the "sigmoid" function.
method considerably improves prediction accuracy. 1
𝑓(𝑥) =
Datasets 1 + 𝑒 −𝑥
Patient data sets, on the other hand, are required to K-Nearest Neighbour
improve organizational research. McDonald’s is in K-Nearest Neighbor (k-NN) allocates a case to the
talks with medical insurance firms about gaining class that is most common among its k closest
access to relevant databases. Researchers can use this neighbours. Distance functions are used to calculate
data to come up with new strategies to save medical the distance between the case and its neighbour.
expenses, save time, and enhance patient
outcomes."At the moment, a quarter of patients will 𝐷
𝐸𝑢𝑐𝑙𝑖𝑑𝑒𝑎𝑛 = √∑𝑘
𝑖=1(𝑥𝑖 −𝑦𝑖 )
2

not react to the medicine," McDonald said, "but we

won't know that until after six weeks of
chemotherapy." "We aim to have tools shortly that Support Vector Machine
can properly forecast a patient's odds of responding to
first-line treatment, and if they don't, we can reliably SVM (Support Vector Machine): Based on the
predict the next medicine to try." support vectors, the Support Vector Machine finds an
ideal hyperplane that optimally divides the classes. overfitting, the majority of the information in the
The role of the kernel in SVM is to accept data and features is reserved, and the redundant information is
transform it into the desired format. In the SVM deleted.
model, the kernel function is a linear function.
3. Deep Learning Model
𝑘(𝑥𝑖 , 𝑥𝑗 ) = 𝑎 < 𝑥𝑖 , 𝑥𝑗 > + 𝑏
Nine neurons make up the first hidden layer, which is
In this work, the kernel SVM is an SVM technique followed by a GaussianNoise layer to boost
that utilises a Gaussian radial basis function (RBF) as robustness and a dropout layer to decrease overfitting.
a kernel. The procedure is then repeated, with one neuron in
the final thick layer. The function of activation is relu.
Naïve Bayes
In a Bayesian environment, the NB algorithm uses the
Maximum A Posteriori decision rule to determine
classifications.
𝑛

𝑦 = 𝑎𝑟𝑔𝑚𝑎𝑥𝑐𝑖 (𝑝(𝑐𝑖 ) ∏ 𝑝(𝑥𝑗 |𝑐𝑖 ))

𝑗=1

Decision Tree
The decision tree is displayed in the form of a tree.
The test items are categorised according to their
feature values. In a decision tree, a node represents an
instance, a branch represents the test results, and the
leaf node represents the class label.
Random Forest
A random forest is a collection of separate decision Experimental Section
trees. Each decision tree produces a classification
During the experimental section following were the
prediction. It determines the test object's class by
steps included.
combining votes from several decision trees.
1. Importing libraries
2. Feature Selection
When the amount of features in a machine learning
method is redundant, the accuracy drops [16]. We
used feature selection on the data to increase the
accuracy of the models and minimise overfitting. To
choose characteristics from the dataset for the
conventional models, we employed two strategies.
We produce the feature significance of the latest
training result for the Decision Tree and Random
Forest models and pick features accordingly.
Correlation depicts how the dataset's characteristics
2. Exploratory Data Analysis (EDA)
are connected to one another. It's easy to see which
a. Providing file path
features are significantly connected when you use the
heatmap view. We may plot the heatmap for a better
look using the seaborn library. We pick only one
feature to represent all of the characteristics in each
set of strongly associated features. To avoid
 b. Reading data from the file 9. Getting the correlation
10. Training of the dataset
logistic regression accuracy: 0.9912087912087912
Decision tree accuracy: 1.0
Random forest accuracy: 0.9978021978021978
11. Compiling the data

569 rows × 32 columns

c. The dataset has 569 rows and 33 columns.
There are two extra columns “id” and
“Unnamed: 32”. We drop Unnamed: 32
which has all Nan values.
d. breast_cancer.info()
e. breast_cancer.info()
3. Counting the values of benign and malignant
from dataset
B 357
M 212
Name: diagnosis, dtype: int64
4. Showing the counts of observations in each
categorical bin using bars
5. Converting the value of malignant and benign 12. Fitting the model
into 1 and 0 13. Diagnosis of the cancer
M=1
B=0
6. Making pairplot by using seaborn of diagnosis

7. Making heatmap of correlation

8. Testing of the dataset
14. Checking of Accuracy training and testing, key features are chosen, with our
criterion for feature selection ranging from selecting
solely mean value attributes to selecting
characteristics with a good correlation with the
predictions. The next stage is to categorise our results
using various classification methods into two
categories: benign and malignant. Finally, we
calculate and examine the performance of these
models in a variety of scenarios.
We will compute the accuracy score of each model in
the experiment. For the test dataset, it compares the
genuine value of diagnosis with the corresponding
predicted value of diagnosis. After that, we'll see the
model's categorization report, which will be the most
accurate. Now we have 357 benign and 212 malignant
values in our dataset. Our dataset is divided into 75
percent training and 25 percent testing, including 267
benign and 159 malignant cells in the training set and
90 benign and 53 malignant cells in the test set.
Correlation is a mathematical term that indicates the
degree of dependence between two variables. It's also
used in data analysis and machine learning to see how
the qualities and the label are related.
Discussion of the results
We utilised Seaborm and Matplotlib visualisation to
As previously stated, we will pre-process the data to visualise our findings. With a 60 percent -40 percent
improve prediction accuracy. UCI Machine Learning training-testing split (838/560), we used our models
Repository provided the data (Wisconsin dataset). on the improved Breast Cancer Wisconsin dataset.
Our data is now raw, and we must pre-process it to Additional training-testing portions of 70% – 30%
convert numeric values to nominal values. Then, for and 80% – 20% were used for conventional
approaches. The number of neighbours was set to 5
for the k-nearest neighbour model, and the
"Minkowski" distance function was utilised. Linear
kernel was utilised for the SVM model. For kernel
SVM, RBF was used as the kernel function. The
quality of a split was measured using the entropy
function in the decision tree model. The number of
estimators in the random forest model was set to ten,
and the entropy function was used to assess split
quality.
The accuracy of logisctic regression and SVM
reduces as the fraction of the training set increases.
The LR model performed an excellent job at
identifying benign instances, however there are a lot
of malignant cases that were misclassified. The SVM
model performed somewhat better in categorising
malignant instances than the LR model, but 3. “Breast cancer misdiagnosis and mammography
performed slightly worse in classifying benign ones. errors.”
On k-NN, Kernel SVM, and NB models, changing the https://hackmd.io/@shaochia/Hk6fwgI1r?type=
training set ratio had no effect. The NB and Kernel view1Introduction, 2011.
SVM models classified malignant cases with 4. L. Wang, “Early diagnosis of breast cancer,”
excellent accuracy, however the k-NN model did not. Sensors, vol. 17, no. 7, p. 1572, 2017.
However, the two models' benign case 5. K. R. Foster, R. Koprowski, and J. D. Skufca,
misclassification rates were considerable, whereas “Machine learning, medical diagnosis, and
the kNN model's was very low. biomedical engineering research-commentary,”
Biomedical engineering online, vol. 13, no. 1, p.
We loaded the libraries before performing
94, 2014.
exploratory data analysis throughout the execution of
6. Chaurasia and S. Pal, “Data mining techniques:
this project. After that, we read data from a csv file.
to predict and resolve breast cancer
There are 569 rows and 33 columns in the dataset.
survivability,” International Journal of
"id" and "Unnamed: 32" are two more columns. We
Computer Science and Mobile Computing
remove Unnamed: 32 from the equation because it
IJCSMC, vol. 3, no. 1, pp. 10–22, 2014.
contains all Nan values. After that, we run a command
7. K. Dubey, U. Gupta, and S. Jain, “Analysis of
to count the benign and malignant values in the
kmeans clustering approach on the breast cancer
dataset. I used seaborn to plot these data on the graph
wisconsin dataset,” International journal of
once I discovered them. The data was then tested,
computer assisted radiology and surgery, vol. 11,
trained, and collated. Following that, we matched the
no. 11, pp. 2033–2047, 2016.
model and diagnosed cancer. Finally, we double-
8. D. Lavanya and K. U. Rani, “Ensemble decision
checked the model's correctness.
tree classifier for breast cancer data,”
Conclusion International Journal of Information Technology
Random Forest received a high score of 0.97 in terms Convergence and Services, vol. 2, no. 1, p. 17,
of accuracy, disregarding PCA. Logistic regression 2012.
(0.923) and K-Neighbors (0.9349) aren't far behind. 9. X.WangandO.Gotoh,“Arobustgeneselection
For accuracy, SVM receives a score of 0.917. method for microarray-based cancer
Decision Tree performs the worst of the six, resulting classification,” Cancer informatics, vol. 9, pp.
in a score of 0.834. Except for the Decision tree, the CIN–S3794, 2010.
application of PCA negates the accuracy of all 10. R. Marshall, “Artificial neural networks in
methods. The accuracy figures, however, are still cancer management,”
greater than Random Forest's. Also, do extremely 11. Purwar and S. K. Singh, “Hybrid prediction
well after using PCA, even if there is a drop in model with missing value imputation for medical
accuracy (0.917). Much may be determined about data,” Expert Systems with Applications, vol. 42,
algorithm performance when additional performance no. 13, pp. 56215631, 2015.
matrices are taken into consideration. Without PCA 12. J. Dheeba, N. A. Singh, and S. T. Selvi,
presentation, decision trees and K-neighbors perform “Computeraided detection of breast cancer on
better, but Logistic Regression and SVM perform mammograms: A swarm intelligence optimized
better once PCA is applied to the data. wavelet neural network approach,” Journal of
biomedical informatics, vol. 49, pp. 45–52, 2014.
References 13. D. Team, “Kernel functions-introduction to svm
1. C. Society, “How common is breast cancer?.” kernel examples.” https://data-
https://www.cancer.org/cancer/breastcancer/ab flair.training/blogs/svm-kernelfunctions/, 2018.
out/how-common-is-breast-cancer.html, 2011. 14. S. B. Ranjit Panigrahi, “Classification and
2. UCHealth, “How accurate are mammograms?.” analysis of facebook metrics dataset using
https://www.uchealth.org/today/how-accurate- supervised classifiers,” Social Network
aremammograms/, 2015. Analytics, 2019.
15. M. B. Kursa, W. R. Rudnicki, et al., “Feature
selection with the boruta package,” J Stat Softw,
vol. 36, no. 11, pp. 1–13, 2010.
16. vishabh goel, “Building a simple machine
learning model on breast cancer data.”
https://towardsdatascience.com/building-a-
simplemachine-learning-model-on-breast-
cancer-dataeca4b3b99fa3, 2018.
17. ”Ultrasound characterisation of breast masses”,
The Indian journal of radiology imaging by S.
Gokhale., Vol. 19, pp. 242-249, 2009. K. Elissa,
“Title of paper if known,” unpublished.
18. “Breast Cancer Prediction Using Genetic
Algorithm Based Ensemble Approach” by
Pragya Chauhan and Amit Swami, 18 October
2018
19. “On Breast Cancer Detection: An Application of
Machine Learning Algorithms on the Wisconsin
Diagnostic Dataset” by Abien Fred M. Agarap, 7
February 2019
20. “Analysis of Machine Learning Techniques for
Breast Cancer Prediction” by the Priyanka
Gupta and Prof. Shalini L of VIT university,
vellore, 5 May 2018.
21. “Breast Cancer Diagnosis by Dierent Machine
Learning Methods Using Blood Analysis Data“
by the Muhammet Fatih Aslan, Yunus Celik ,
Kadir Sabanci and Akif Durdu, 31 December,
2018
22. ”Performance Evaluation of Machine Learning
Methods for Breast Cancer Prediction”, by
Yixuan Li, Zixuan Chen October 18, 2018
23. “Breast Cancer Prediction and Detection Using
Data Mining Classification Algorithms: A
Comparative Study” by Mumine Kaya Keles, Feb
2019
24. “Breast Cancer Prediction Using Data Mining
Method ” by Haifeng Wang and Sang Won Yoon,
Department of Systems Science and Industrial
Engineering State University of New York at
Binghamton Binghamton, May 2015.