Journal of Artificial Intelligence Research & Advances

ISSN: 2395-6720
Volume 7, Issue 1
www.stmjournals.com

Malaria Parasite Detection in Thin Blood Smear Images

Using Deep Learning
Rouf Ahmad Tantray1,*, Zahoor Ahmad Najar2
1
Student, Department of Information Technology, Central University of Kashmir, Jammu and
Kashmir, India
2
Assistant Professor, Department of Information Technology, Central University of Kashmir, Jammu
and Kashmir, India

Abstract
Algorithms in deep neural networks are being used to show the presence of malaria and to
estimate the depth of infection by automatically counting individual uninfected and infected
RBCs in images of thin blood smears. During the training period, the relationship was tried on
a set of 13600 images from several thin blood spreads and experiment was conducted. My
dataset is divided into 80% for training and 20% for testing. The relationship between the
results from the algorithm and expert human readers was r = 0.866. Using image analysis
results, the level of parasitemia may be achieved by applying methods to images of thin film
smears. In this work, convolutional neural networks like CNN and ResNet are used for
classification purposes and then classify them either infected or uninfected. Since my dataset
contains only two prevalent species like Plasmodium vivax and P. falciparum that are
predominant in their different cell stages and can cause 99% of the total infections. The results
indicate varied distribution prevalence of P. vivax and P. falciparum, and also the mixed species
infection due to these two species.

Keyword: Malaria parasite, deep learning, computer vision, intelligent systems, convolutional
neural network.

*Author for Correspondence: E-mail: [email protected]

INTRODUCTION The World Malaria Report 2017 [1] tells us the

Transmission of malaria is caused by female fact that the progress against the malaria has
anopheles mosquito bites. Then it gets mixed been stalled in several countries and that the
up with the blood stream through blood world was unlikely to achieve the WHO Global
transfusion. The tropical and subtropical area technical strategy for malaria 2016–2030. Our
particularly in the African region makes main goal must be to have a proper strategy
malaria a dreadful disease and causes several including efforts to intensify the response in the
million deaths in those regions. One of the highest burden countries.
reasons is that the temperature is optimal and
the breeding can takes place faster as compared This study related to check the intensity of
to other regions. The World Health parasitemia inside the infected cell aims to
Organization’s (WHO) [1] 11th World Malaria build one such automated system—an
Report summarizes global progress in the fight intelligent system that can detect the presence
against malaria up to the end of 2017. Every of malaria parasites and identify amongst the
year people in millions die due to malaria specific species present in blood samples under
disease. Some reports suggest that there were observation.
about 214 million cases of malaria globally in
2015 and about 438000 malaria deaths. The The parasitemia of Plasmodium vivax is not too
sub-Saharan African countries are the worst much heavy. The patients’ blood samples
sufferers. This is because of the environmental which are infected erythrocytes contain only
conditions of some Africa regions that are one trophozoite. In this trophozoite stage, the
favorable for mosquito's breeding. structure of the parasite is well-defined, having

JoAIRA (2020) 28-32 © STM Journals 2020. All Rights Reserved Page 28
Malaria Parasite Detection in Thin Blood Smear Images Using Deep Learning Tantray and Najar

a prominent central vacuole, whereas eroded to highlight the parasites and platelets
falciparum is most recognizable having sickle- (foreground). The cells and parasites are then
shaped gametocyte. P. falciparum is considered identified based on a combination of annular
the most severe threat among all the species of ring ratio method, size, and intensity variation.
Plasmodium vector, posing high rate of In conjunction with the techniques explored in
complications and higher death rates every year the previously mentioned researches, machine
including in India. learning techniques serve as the core for
decision systems for malaria. Some like Das et
RELATED WORK AND al. [3], employ generative methods such as
LITERATURE SURVEY multivariate regression models. Anggraini et al.
Detecting malaria parasite images of thin blood [2] used Bayes decision theory to classify
smears, it is required to have a proper images after segmentation. Other researchers
knowledge of observing the image. Using use discriminative models to classify malaria
proper image analysis techniques, one can infer images.
from the images the detailed FoVs (Field of
Views) and corresponding to them also exists D. Bibin et al. [5]—malaria parasite detection
the various types of features like color, rotation from peripheral blood smear images using DBN
and pixel density values. The learning rate is studies in the domain of deep learning indicate
determined through the use of various neural that there is a profound increase in the use of
network techniques. The diagnosis of malaria is deep belief networks (DBNs) due to the
done through demonstration of malaria advancements in the effective layer-by-layer
parasites in the blood by preparing blood learning method. DBNs have been used in
smears. A thin smear is used in determining the various machine learning areas, such as image
infecting species. Such smears are usually analysis, speech recognition, information
retrieval, voice activity detection (VAD) and
prepared whenever blood is drawn from a
natural language understanding. DBN-related
fingertip, applied to a slide, and stained by one
models are also used for 3-D object recognition,
of the stains such as Giemsa to highlight the
hand-written character recognition, motion
parasite. A similar research was done by capture data modeling, information retrieval, and
Anggraini et al. [2]. The proposed algorithm machine transliteration. Compared to the
does not consider for segmentation the existing studies in the classification of malaria
individual pixels of the image. Instead, a global parasites, this study contributes by providing a
threshold was obtained by varying the contrast more efficient and simplified approach. The first
among pixels, and was used to classify each article to apply deep learning to malaria
pixel as belonging to either foreground or diagnosis is by Liang et al. [6], who use a
background. To do this, pre-processing of the convolutional neural network to discriminate
images to obtain “uniform” images was done by between infected and uninfected cells in thin
converting all input images to gray scale. The blood smear images after applying a
images were then filtered to normalize the pixel conventional level-set cell segmentation
intensities around a median value before approach. This is an ideal application for deep
obtaining the histogram of the individual learning because images of segmented RBCs are
images and expanding each until the two a natural input for a convolutional neural
intensity classes, foreground and background, network.
become distinct given threshold intensity.
Some researches employ more morphological Gopakumar et al. [7] suggests CNN for image
techniques than the two previous ones. The analysis that operate directly on focus stack of
work done by Das et al. [3] uses thresholding, images is employed to identify malarial
marker-controlled watershed algorithm, and infection and compare the results of detection
Haralick textural features extraction. Another of malaria-infected RBCs in terms of sensitivity
work by Kareem et al. [4] also uses and specificity obtained by a support vector
morphological image transformations to infer machine classifier trained on the statistical and
malaria parasites from Giemsa-stained blood textural features extracted from the suspected
films. First, a grayscale image is dilated and parasite locations.

JoAIRA (2020) 28-32 © STM Journals 2020. All Rights Reserved Page 29
Journal of Artificial Intelligence Research & Advances
Volume 7, Issue 1
ISSN: 2395-6720

Hung et al. [8] choose to use a deep learning RGB candidate patch which is obtained from
based framework called faster region-based the best focused image having higher variance
convolutional neural network (Faster R-CNN) measure (Figure 1).
because all the previous attempts were
particularly made in order to automate the Dealing with patches around the suspected
process of identifying and quantifying malaria parasite locations, the input to CNN is decided
using complex workflows for image processing as 32*32 which is decent enough to hold the
and machine learning classification such as neighborhood in making the final decision. We
features from a predetermined set of have chosen standard kernel size 7*7 for the
measurements (intensity, shape, and texture) feature extraction. Being a binary classification,
have been the basis of the top performing object the output neurons were set to 2 that can be
detection models in recent years. In contrast to either infected or uninfected. Once the input
previous methods, this one avoids the task of and output are set, a reasonably deep CNN must
segmentation and does not rely on general give a fair classification.
features for classification.
Feature Extraction and Classification
TWO BEST APPROACHES For each 300×300 pixel input image location,
Deep Learning 50×50 pixel patches are extracted with each
The full-sized images which are used for individual cell image. The pixel region displays
training the deep learning models, we select a pixel information: (x, y) pixel intensity value
patch around the point of local minimum in from set of blood images. The distance measure
range of 50×50 crops of each full image. This of the object with respect to the axis x data and
is done in order to augment the number of y data is determined with angular distance.
training examples and cut down on training With the help of this angular distance, the
time. Enough crops of each full image were number of pixels is calculated using image
taken such that the density of cells contained in analysis classifying the malaria as either non-
each crop will be at least twice the number of parasite RBC and parasite. The third stage of
cells contained in the full-sized image up to a three-stage object detection procedure involves
maximum of 100 crops [9]. Also, there will be Softmax classifier which generalizes the
some imbalance in the cell count where the logistic regression classifier. Softmax classifier
training data is highly imbalanced and (SMC) is added to effectively differentiating
underrepresented towards RBCs, in order to malaria-infected and noninfected blood smear
create a more balanced training set, we have images. As the model accepts high-level data as
rotated each crop by 90 degrees, which is a kind input and reconstructs an output of better clarity
of augmentation where the cell count of and accuracy than the input [10]. This
underrepresented cell roughly four times and reconstructed output maps between the input
removed crops containing only RBCs. The deep and the class labels fed to the supervised model
neural network like CNN does not depend on of SMC produces a practical solution of value
hand crafted features, it works on custom build 0 and 1, that can be interpreted as malaria-
models where it learns while training data and infected (abnormal) or noninfected (normal)
predicts accordingly. It directly operates on thin blood smear image (Figure 2).

Fig. 1: Flowchart Showing Detection of Malaria Parasite Locations.

JoAIRA (2020) 28-32 © STM Journals 2020. All Rights Reserved Page 30
Malaria Parasite Detection in Thin Blood Smear Images Using Deep Learning Tantray and Najar

Fig. 2: Three-stage Image Detection Workflow Object Detection Method.

CONCLUSIONS able to provide quantification and exact

Conventionally, the accuracy of malaria locations of uninfected and infected blood
diagnosis depends on the human expert. We images. Such a system would either enable a
will propose an automated diagnosis system fully automated diagnosis or act as an aid to
using deep learning principles that will improve accuracy and lower workload of
exclude the human expert from the process or human experts. In this paper, we have
serve as an aid for an expert to lower workload proposed an automated state of the art machine
and improve accuracy. To propose an which works on custom-built features,
algorithm based on CNN that can deal with the portable, cost-effective prototype system with
problems inherent in low quality images. necessary instrumentation. Image
analysis/classification algorithms that can be
To conclude my contributions, we have used to quantify malaria parasitemia. The
investigated and proposed a solution to the results produced in this paper suggests that
diagnosis of malaria in thin blood smear implementing CNN over traditional classifiers
images. Also splitting the problem into two like SVM gives considerable improvement in
parts, firstly localization of erythrocytes, and terms of higher specificity and higher
secondly, the classification of these as sensitivity and also suggest operating on focus
uninfected or infected. The system must be stack of images for malaria diagnosis.

JoAIRA (2020) 28-32 © STM Journals 2020. All Rights Reserved Page 31
Journal of Artificial Intelligence Research & Advances
Volume 7, Issue 1
ISSN: 2395-6720

REFERENCES analysis for malaria diagnosis. 2016 IEEE

1. WHO. Malaria microscopy quality International Conference on
assurance manual-version 2. World Health Bioinformatics and Biomedicine 1(BIBM),
Organization; 2016. World Malaria Report 15–18 Dec. 2016, Shenzhen, China, pp.
2016. World Health Organization; 2016. 3. 493–496.
Tek FB, Dempste. 7. Gopakumar G, Jagannadh VK, Gorthi SS,
2. Anggraini D, Nugroho A, Pratama C, Rozi Subrahmanyam GRKS. Framework for
I, Iskandar A, Hartono R. Automated status morphometric classification of cells in
identification of microscopic images imaging flow cytometry. Journal of
obtained from malaria thin blood smears. Microscopy. 2016;261(3):307–319p.
Proceedings of the 2011 International 8. Ren S, He K, Girshick R, Sun J. Faster R-
Conference on Electrical Engineering and CNN: Towards real-time object detection
Informatics, Bandung, Indonesia. 17–19 with region proposal networks. IEEE
July 2011. Transactions on Pattern Analysis and
3. Das D, Ghosh M, Chakraborty C, Maiti A,
Machine Intelligence. 2017;39(6):1137–
Pal M. Probabilistic prediction of malaria
1149p.
using morphological and textural
information. Proceedings of the 2011 9. Russakovsky O, Deng J, Su H, Krause J,
International Conference on Image Satheesh S, Ma S, Huang Z, Karpathy A,
Information Processing (ICIIP), Shimla, Khosla A, Bernstein M, Berg AC, Fei-Fei
India. 3–5 November 2011; pp. 1–6. L. ImageNet large scale visual recognition
4. Kareem S, Kale I, Morling R. Automated challenge. International Journal of
Malaria Parasite Detection in Thin Blood Computer Vision. 2015;115:211–252p.
Films: A Hybrid Illumination and Color 10. Alex Krizhevsky, Ilya Sutskever, Hinton
Constancy Insensitive, Morphological GE. ImageNet classification with deep
Approach; Proceedings of the 2012 IEEE convolutional neural networks.
Asia Pacific Conference on Circuits and Communications of the ACM.
Systems (APCCAS); Kaohsiung, Taiwan. 2017;60(6):84–90p.
2–5 December 2012.
5. Bibin D, Nair MS, Punitha P. Malaria
parasite detection from peripheral blood Cite this Article
smear images using deep belief networks. Rouf Ahmad Tantray, Zahoor Ahmad
IEEE Access. 2017;5:9099–9108p. Najar. Malaria Parasite Detection in Thin
6. Liang Z, A Powell, Ersoy I, Poostchi M, Blood Smear Images Using Deep Learning.
Silamut K, Palaniappan K, Guo P, Hossain Journal of Artificial Intelligence Research
MA, Sameer A, Maude RJ, Huang JX, & Advances. 2020; 7(1): 28–32p.
Jaeger S, Thoma G. CNN-based image

JoAIRA (2020) 28-32 © STM Journals 2020. All Rights Reserved Page 32