DOI: 10.25122/jml-2019-0090 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp.

382–387

Deep Learning for Acute Myeloid Leukemia Diagnosis

Elham Nazari1, Amir Hossein Farzin2, Mehran Aghemiri3, Amir Avan4, Mahmood Tara1, Hamed Tabesh1*
1. Department of Medical Informatics, Faculty of Medicine, Mashhad University of Medical Sciences, Mashhad, Iran
2. Department of Computer Engineering, Khayyam University, Mashhad, Iran
3. Department of Medical Informatics, Faculty of Medical Sciences, Tarbiat Modares University, Tehran, Iran
4. Molecular Medicine Group, Department of Modern Sciences and Technologies, School of Medicine, Mashhad University of
Medical Sciences, Mashhad, Iran

* Correesponding Author:
Hamed Tabesh
Department of Medical Informatics, Faculty of Medicine,
Mashhad University of Medical Sciences, Mashhad, Iran
Phone: +98 51 38002536;
Fax: +98 51 38002445.
E-mail: [email protected]

Received: September 8th, 2019 – Accepted: December 29th, 2019

Abstract
By changing the lifestyle and increasing the cancer incidence, accurate diagnosis becomes a significant medical action. Today, DNA
microarray is widely used in cancer diagnosis and screening since it is able to measure gene expression levels. Analyzing them by
using common statistical methods is not suitable because of the high gene expression data dimensions. So, this study aims to use new
techniques to diagnose acute myeloid leukemia.
In this study, the leukemia microarray gene data, contenting 22283 genes, was extracted from the Gene Expression Omnibus repos-
itory. Initial preprocessing was applied by using a normalization test and principal component analysis in Python. Then DNNs neural
network designed and implemented to the data and finally results cross-validated by classifiers.
The normalization test was significant (P>0.05) and the results show the PCA gene segregation potential and independence of cancer
and healthy cells. The results accuracy for single-layer neural network and DNNs deep learning network with three hidden layers are
63.33 and 96.67, respectively.
Using new methods such as deep learning can improve diagnosis accuracy and performance compared to the old methods. It is rec-
ommended to use these methods in cancer diagnosis and effective gene selection in various types of cancer.

Keywords: AML, machine learning, deep learning, neural network, microarray.

Introduction complications such as infections and bleeding [4]. There-

fore, microarrays improve clinical diagnosis providing high
accuracy diagnostic procedures. It can be used as a gold
Big medical data generated as a result of recent advanc- standard to diagnose [5] and help to treatment progress
es in biology. Using appropriate analyzing methods led development and understanding cell biology, especially in
biologists to percept the complex dynamic system of life. oncology studies. Thus, the gene expression patterns are
It is one of the most critical challenges for biologists. Mi- compared in two healthy and cancerous tissues. Cancer is
croarray is well known despite it is a new technology in closely linked to genetic changes [6]; hence, the accurate
molecular biology. It is used to monitoring genome-wide cancer diagnosis speeds up by using a pattern to classify
expression levels by biologists [1]. This technology in- normal and cancer cells [7]. So, a timely and accurate di-
cludes examining a thousand genes and protein activity on agnosis is essential.
a small scale to compare the similarity and track changes There are many pieces of evidence that accurate can-
such as track decreasing or increasing gene activities and cer diagnosis is one of the most effective ways of reducing
track protein samples changes in comparison to the control the mortality rate [7,8]. The higher number of genes (p)
sample [2]. Microarray uses include genotyping, epigenet- compare to the number of tissues (n) is the feature of the
ics, translation profiling, gene expression profiling [3]. Mi- microarray [9]. It is challenging to analyze microarray data
croarray can be replaced by aggressive cancer detection using statistical methods to classify high-dimensional data
methods such as bone marrow biopsy, which is used in (p>n) due to overestimation and multiple linearity prob-
the accurate diagnosis of acute myeloid leukemia. Bone lems make statistical classification of microarrays difficult
marrow biopsy is invasive, painful, and can cause serious [10, 11]. Its analyzing methods are evolving rapidly, and

382
 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp. 382–387

there is no specific way considered the best way to ana- layer gets data (dendrite), the hidden layer processes data
lyze microarrays [2]. Recently, expert systems to diagnose (soma and axon), and finally, processed data is sent to the
cancerous gene data are increasing, and machine learn- output layer (synapse)(Figure 2) [18-21].
ing techniques are currently used more. Machine learning
can help to automation and intelligence process, improve
development, accuracy and reducing costs [12]. Machine
learning, ensemble methods, and deep learning are show-
ing high performance in classifying biological data [13-16].
In this study, neural networks and deep learning were
used to separate healthy and cancerous cells in leukemia
related genes. Acute myeloid leukemia (AML) is the type
of cancer that starts in the bone marrow, but in most of the
cases, it moves to the blood very fast. This type of cancer
worsens fast if left untreated [17].

Figure 2: Structure of an artificial neural network.

Material and Methods

In this study, we classified healthy and cancerous cells by The neural network’s behavior is shaped by the architec-
neural networks and deep learning. ture of that network. Neural network architecture can be
defined as follows:
Artificial neural network • The number of neurons;
• Number of layers;
It is a computational and algorithmic model inspired by the • Types of communication between layers.
structure and functional aspects of biological neural net- Perceptron is one of the simplest neural networks. It is
works and the concept of neurons. It is considered one of a learning algorithm for a binary classifier, called a thresh-
the nonlinear statistical data modeling tools and is used old function:
for pattern recognition and modeling complex relations
between inputs and outputs. It consists of some simple
units that work in parallel. Weighting between units is the
primary way to store information long-term and learn new
information by updating weights. w: vector of real value weights; b: bias
A neuron of the human nervous system consists of
dendrites, a single axon, soma, and nucleus, as shown in
Figure 1. The main configuration of perceptron networks is shown
in Figure 3.
Activation functions are used to propagate the node
outputs from one layer to the next (up to the output layer).
The activation function is a function that activates the neu-
ron. There are several types of activation functions such as
Identity, Binary Step, Sigmoid, Tanh, ReLU, Leaky ReLU,
and Softmax.
Sigmoid is a widely used activation function convert-
ing illimitable independent variables to simple probabilities
between 0 and 1. Sigmoid can infinitely reduce data or
outlying values without deleting them. Unlike the sigmoid
activation function, Tanh is bound to the (-1,1) range. It is
worth mentioning that tanh deals easier with negative num-
bers. Besides, tanh is a well-liked and widely used activa-
tion function. The Softmax function is a multiclass logistic
regression and a generalization of the sigmoid. Therefore,
Figure 1: Structure of a typical neuron. it can be applied to continuous data (rather than binary
classification). Rectified linear units (ReLU) are based on
the latest scientific advances, and it has been proven to
Dendrite receives electrochemical impulses from the other be working in many conditions. ReLU is bound to the [0,
neurons. The soma processes these signals. The output inf) range. It makes the network lighter and efficient due
is transmitted to terminal dendrites by axons, where these to the characteristics of ReLU. Also, these activation func-
new impulses are sent to the next neuron. An artificial neu- tions show better performance than Sigmoid in training
ral network works the same way on three layers: the input data. Recent studies show that deep learning networks

383
 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp. 382–387

Figure 3: A single-layer perceptron.

using ReLU are able to train well without preprocessing networks (RNNs), convolutional neural networks (CNNs)
techniques. and more [19].
Loss functions determine how much a trained neural
network is close to reality. It is a measured inconsisten- Quality control
cy between the real and predicted value considered as
an error; the mean error value takes and represents the This is one of the steps of the microarray data analyzing,
difference between the real-world and the neural network. after which it is possible to test and interpret the method.
There are many common loss functions, such as MAE, Any negligence to impose quality control may cause
MAPE, MSLE, and MSE [19]. detour and alter the results significantly for many reasons,
such as the following:
Learning rate 1. The biologist grows the cell culture without know-
ing that bacteria may live in the cell.
It determines the neural network values and how much 2. There may be fungal or viral contamination.
they change by new training data. The learning rate is set 3. The RNA treatment may not do well after RNA ex-
before the learning process begins. A low learning rate traction.
means more time to train, but a high learning rate makes 4. Because RNA is a highly unstable molecule and it
the network more sensitive to new information [20, 21]. begins to crumble, the quality decreases at room
temperature.
Deep learning 5. The sample size is not enough, or there is an error
in complementary DNA (cDNA) generation in rank
Recently, a new machine learning technique known as steps.
deep learning, is used frequently. New studies show that 6. The results are not reliable if something goes
this algorithm has better results compared to machine wrong during the scanning or hybridization steps.
learning, for example, identifying and discovering drugs, Biases that occur in the study results related to genetic
image processing, and speech [22-27]. data lead to false-positive and false-negative results.
Deep-learning is defined as a neural network with a The genes that can separate the cancerous and healthy
large number of parameters and layers. In fact, it is a class cells indicate that the experiment is well done. So, dimen-
of machine learning algorithms that uses a hierarchical sion reduction techniques are used to detect important
nonlinear structure in multiple layers to extract features genes in separating these samples. Otherwise, co-expres-
and transformations [19]. sion or co-relation between genes or between samples can
Unlike other machine learning methods requiring an be measured. Actually, the purpose of dimension reduction
expert to extract features, deep learning can act as an is to capture the variations in microarray data [30, 31].
automatic feature extractor that transforms low-level fea-
tures into higher-level abstractions [28]. In addition, deep PCA
learning can incorporate momentary, indirect and minor
changes and leads to higher accuracy than other machine Principal component analysis (PCA) is an analysis of sim-
learning methods [29]. plifying high-dimensional complexity, including patterns
Types of deep learning techniques can include deep and trends. High-dimensional data is common in biology,
neural networks (DNNs), autoencoders networks (AEs), and multiple features occur when the expression of the dif-
generative adversarial networks (GANs), repeating neural ferent genes for each sample is measured [32, 33].

384
 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp. 382–387

Dataset TensorFlow is an open-source library for large-scale nu-

In this study, we used the Gene Expression Omnibus merical computing and machine learning techniques, mak-
(GEO) database and AML-related data, and healthy and ing it easier and faster to apply machine learning and deep
cancerous cases were extracted. GEO is a public and in- learning techniques.
ternational database that publishes free genomic data ob- Pandas provides fast, flexible and illustrative data
tained from microarray studies, Next Generation Sequenc- structures in Python. It also enables data entry, manipula-
ing (NGS), sequence-based functional genomics studies, tion, and data analysis.
and handles high-throughput data submissions. Numpy is an open-source package that supports sci-
Data from 36 cases containing 22283 gene expression entific calculations, matrices, and multidimensional arrays.
of AML microarray experiments were retrieved from the It also supports functions such as Sine, Cose, Log, and
GSE9476 dataset, including 10 healthy and 26 AML cases. others.
The dataset was registered in 2007 and updated in 2017. Sklearn is an open-source package in Python that of-
fers powerful data analysis and data mining tools.
Data preparation
Analysis of results related to AML data
Microarray data normalization is one of the essential mi-
croarray data analyzing steps, balancing the hybridization The PCA result is shown in Figure 4.
intensity of each point in the data matrix. Comparability of As can be seen, the cancerous and healthy data were
samples was examined by Python 3.7 and data normality separated, indicating that the experiment was performed
was confirmed [34]. correctly and had the prerequisite to implement the meth-
In this study, the PCA method was used for dependen- od, and the results are reliable. The result of a simple neu-
cy testing and data quality. ral network and the deep learning-based network imple-
mentation is shown below:
A simple neural network with a middle layer has an ac-
curacy of 0.6333, and a deep learning network with three
Results hidden layers has an accuracy of 0.9667
According to the results, neural networks based on
deep learning techniques provided higher accuracy than a
Modeling simple neural network.

Normalized data modeled by TensorFlow, Pandas, Numpy,

Sklearn packages, and activity functions include Softmax,
Adamax, and RELU. These packages explained the fol-
Discussion
lowing.
In this study, we used three hidden layers of DNNs to Microarray technology has enabled thousands of genes
analyze the data. Due to using dropout, the overfitting rate to be analyzed simultaneously and, very important, in the
was reduced, the Adamax function was used as an optim- early detection of diseases, including cancer. It is difficult to
izer, and sparse categorical was used to calculate the loss. analyze these types of data by usual techniques because
Accuracy was considered as the network improvement cri- of their high dimensionality. Therefore, in this study, deep
terion at each epoch. Also, 70% of data was used as train- learning-based techniques and its comparing simple neu-
ing data, and the number of epochs was considered 1000. ral network were used to detect AML. The reported result

Figure 4: PCA for microarray data.

385
 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp. 382–387

shows higher accuracy in the classification of cancerous References

and healthy tissues by a deep learning-based network. 1. Babu MM. Introduction to microarray data analysis. Computational
Deep learning methods are a class of machine learn- genomics: Theory and application. 2004;225:249.
ing techniques able to identify patterns with high complex- 2. Akay M-F. Support vector machines combined with feature selec-
ity in large datasets, and it can analyze data with large tion for breast cancer diagnosis. Expert Syst Appl. 2009;36:3240-7.
numbers of features and samples with high accuracy. 3. Shaik NA, Hakeem KR, Banaganapalli B, Elango R, editors. Es-
The AML dataset had 22283 features, so it was used for sentials of Bioinformatics, Volume I: Understanding Bioinformatics:
modeling by deep learning [35, 36]. Today, this is a state Genes to Proteins. Springer; 2019.
of the art of machine learning techniques. The high per- 4. PERCIVAL, Mary-Elizabeth, et al. Bone marrow evaluation for di-
agnosis and monitoring of acute myeloid leukemia. Blood reviews,
formance of these types of machine learning methods
2017, 31.4: 185-192.
in various industries, including healthcare, medicine and 5. Yoo S, Choi J, Lee S, Yoo N. Applications of DNA microarray in
bioinformatics, has been confirmed [35]. This technique disease diagnostics. J Microbiol Biotechnol. 2008;19(7): 635–46.
presents valuable results in various medical fields such as 6. Sack G-H. Medical Genetics. United States: McGraw-Hill Compa-
thyroid diagnosis using images [37], heart disease detec- nies; 1999.
tion [38], breast cancer diagnosis [39, 40], molecular prop- 7. Tarver, Talicia. Cancer facts & figures 2012. American cancer soci-
erties identification of drugs damaging the liver [41], orally ety (ACS) Atlanta, GA: American Cancer Society, 2012. 66 p., pdf.
disintegrating tablets (ODT) formulation prediction using Available from. 2012.
an artificial neural network [42], and predicting water sol- 8. Curry, Susan J., et al. Potential of screening to reduce the burden
of cancer. In: Fulfilling the Potential of Cancer Prevention and Ear-
ubility of medication using other machine learning meth-
ly Detection. National Academies Press (US), 2003.
ods [43]. In addition, it has been exploited in modeling the 9. Kalina J. Classification methods for high-dimensional genetic data.
sequence specificity of DNA–protein binding [44], genes Biocybern Biomed Eng. 2014;34:10-8.
prediction [45], motif identification, binding classification 10. Piao Y, Piao M, Park K, Ryu K-H. An ensemble correlation-based
[46], protein binding [47], predicting genomic sequence, gene selection algorithm for cancer classification with gene ex-
and the effects of non-coding variants [48]. Finding genes pression data. Bioinformatics (Oxford, England). 2012;28:3306-15.
is the most crucial research problem in bioinformatics. 11. Chen K, Wang K, Wang K, Angelia M. Applying particle swarm
Researchers have proposed different models for finding optimization-based decision tree classifier for cancer classification
genes in the DNA sequence, but sometimes it does not on gene expression data. Appl Soft Comput. 2014;24:773–80.
12. Rowe RC, RobertsRJ. Artificial intelligence in pharmaceutical
work because of DNA sequence length variety and low
product formulation: knowledge-based and experts ystems. Pharm
accuracy [49]. SciTechnol Today 1998;1:153–9.
Further studies with this aim are recommended in the 13. Cancer is the second leading cause of death globally 2018. Avail-
future. Also, omic data are beneficial for discovering drugs able from: http://www.who.int/news-room/fact-sheets/detail/cancer
and biomarkers. These data are highly variable, high-di- 14. Liu KH, Tong M, Xie ST, Yee Ng VT. Genetic programming based
mensional, and use multiple sources. ensemble system for microarray data classification. Computational
and mathematical methods in medicine. 2015;2015.
15. Tan AC, Gilbert D. Ensemble machine learning on gene expres-
sion data for cancer classification.
Conclusion 16. Yue T, Wang H. Deep learning for genomics: A concise overview.
arXiv preprint arXiv:1802.00810. 2018 Feb 2.
17. Rubnitz JE, Gibson B, Smith FO. Acute myeloid leukemia. Pediat-
Deep neural networks are efficient algorithms using com- ric clinics of North America. 2008 Feb 1;55(1):21-51.
positional layers of neurons to overcome omic data chal- 18. Dey A. Machine learning algorithms: a review. International
lenges [50]. In future studies, deep learning-based tech- Journal of Computer Science and Information Technologies.
niques for medical image processing, genetic data, audio 2016;7(3):1174-9.
data, handwriting recognition for diagnosis and therapies 19. Patterson J, Gibson A. Deep learning: A practitioner’s approach. “
are recommended due to high performance and the ability O’Reilly Media, Inc.”; 2017 Jul 28.
to manage complexity and data variations. 20. V. Sharma, S. Rai, A. Dev, “A Comprehensive Study of Artifi-
cial Neural Networks”,International Journal of Advanced Rea-
search in Computer Science and Software Engineering, ISSN
2277128X,Volume2,Issue 10,October 2012.
Acknowledgments 21. S. B. Hiregoudar, K. Manjunath, K. S. Patil, “A Survey: Research
Summary on Neural Networks”,International Journal od Research
in Engineering and Technology, ISSN:23191163,Volume 03,Spe-
The present study results from a research project approved cial Issue03,page 385-389,May,2014
by the vice chancellery for research of the Mashhad Uni- 22. Lusci A,PollastriG,BaldiP.Deeparchitecturesanddeeplearningin
versity of Medical Sciences (Grant number 961731). chemo informatics :thepredictionofaqueoussolubilityfordrug-like
molecules. J ChemInfModel 2013;53:1563–75.
23. MaJ,SheridanRP,LiawA,DahlGE,SvetnikV.Deepneuralnetsas-
methodforquantitativestructure–activity relationships. J ChemInf
Conflict of Interest Model 2015;55:263–74.
24. Altae-Tran H,RamsundarB,PappuAS,PandeV.Lowdatadrug dis-
covery with one-shot learning. ACS CentSci 2017;3:283–93.
The authors declare that there is no conflict of interest. 25. The Next Era: Deep Learning in Pharmaceutical Research Sean Ekins.

386
 Journal of Medicine and Life Vol. 13, Issue 3, July-September 2020, pp. 382–387

26. Yang Yilong,et al.Deep learning for in vitro prediction of pharma- 40. Abdel-Zaher AM, Eldeib AM. Breast cancer classification using
ceutical formulations . Acta Pharmaceutica Sinica B (2018),https:// deep belief networks. Expert Systems with Applications. 2016 Mar
doi.org/10.1016/j.apsb.2018.09.010 15;46:139-44.
27. Predicting oral disintegrating tablet formulations by neural network 41. Xu Y, Dai Z, Chen F, Gao S, Pei J, Lai L. Deep learning for drug-in-
techniques Run Han, Yilong Yang, Xiaoshan Li, Defang Ouyang duced liver injury. Journal of chemical information and modeling.
28. Schmidhuber J . Deep learning in neural networks: an overview. 2015 Oct 13;55(10):2085-93.
Neural Netw 2015;61:85–117. 42. Han R, Yang Y, Li X, Ouyang D. Predicting oral disintegrating tablet
29. Rost B ,Sander C . Combining evolutionary information and formulations by neural network techniques. Asian Journal of Phar-
neural networks to predict protein secondary structure. Proteins maceutical Sciences. 2018 Jul 1;13(4):336-42.
1994;19:55–72 . 43. Lusci A, Pollastri G, Baldi P. Deep architectures and deep learning
30. Raman T, O’Connor TP, Hackett NR, Wang W, Harvey BG, Attiyeh in chemoinformatics: the prediction of aqueous solubility for drug-
MA, Dang DT, Teater M, Crystal RG. Quality control in microarray like molecules. Journal of chemical information and modeling.
assessment of gene expression in human airway epithelium. BMC 2013 Jul 2;53(7):1563-75.
genomics. 2009 Dec;10(1):493. 44. Zeng H, Edwards MD, Liu G, Gifford DK. Convolutional neural
31. Anderson CA, Pettersson FH, Clarke GM, Cardon LR, Morris AP, network architectures for predicting DNA–protein binding. Bioin-
Zondervan KT. Data quality control in genetic case-control associ- formatics. 2016 Jun 11;32(12):i121-7.
ation studies. Nature protocols. 2010 Sep;5(9):1564. 45. Sree PK, Rao PS, Devi NU. CDLGP: A novel unsupervised classi-
32. Jolliffe I. Principal component analysis. Springer Berlin Heidel- fier using deep learning for gene prediction. In2017 IEEE Interna-
berg; 2011. tional Conference on Power, Control, Signals and Instrumentation
33. Raychaudhuri S, Stuart JM, Altman RB. Principal components Engineering (ICPCSI) 2017 Sep 21 (pp. 2811-2813). IEEE.
analysis to summarize microarray experiments: application to 46. Lanchantin J, Singh R, Lin Z, Qi Y. Deep motif: Visualizing genom-
sporulation time series. InBiocomputing 2000 1999 (pp. 455-466). ic sequence classifications. arXiv preprint arXiv:1605.01133.
34. Quackenbush J. Microarray data normalization and transforma- 2016 May 4.
tion. Nat Genet. 2002;32:496-501 47. Zeng H, Edwards MD, Liu G, Gifford DK. Convolutional neural
35. Min S, Lee B, Yoon S. Deep learning in bioinformatics. Briefings in network architectures for predicting DNA–protein binding. Bioin-
bioinformatics. 2017 Sep 1;18(5):851-69. formatics. 2016 Jun 11;32(12):i121-7.
36. Zou J, Huss M, Abid A, Mohammadi P, Torkamani A, Telenti A. 48. Yue T, Wang H. Deep learning for genomics: A concise overview.
A primer on deep learning in genomics. Nature genetics. 2018 arXiv preprint arXiv:1802.00810. 2018 Feb 2.
Nov 26:1. 49. Sree PK, Rao PS, Devi NU. CDLGP: A novel unsupervised classi-
37. Ma L, Ma C, Liu Y, Wang X. Thyroid diagnosis from SPECT imag- fier using deep learning for gene prediction. In2017 IEEE Interna-
es using convolutional neural network with optimization. Computa- tional Conference on Power, Control, Signals and Instrumentation
tional intelligence and neuroscience. 2019;2019. Engineering (ICPCSI) 2017 Sep 21 (pp. 2811-2813). IEEE.
38. Tomov NS, Tomov S. On Deep Neural Networks for Detecting 50. Mamoshina P, Vieira A, Putin E, Zhavoronkov A. Applications of
Heart Disease. arXiv preprint arXiv:1808.07168. 2018 Aug 22. deep learning in biomedicine. Molecular pharmaceutics. 2016 Mar
39. Mohamed AA, Berg WA, Peng H, Luo Y, Jankowitz RC, Wu S. A 29;13(5):1445-54.
deep learning method for classifying mammographic breast densi-
ty categories. Medical physics. 2018 Jan;45(1):314-21.

387