Sensors International 4 (2023) 100209

Contents lists available at ScienceDirect

Sensors International
journal homepage: www.keaipublishing.com/en/journals/sensors-international

A new ensemble learning approach to detect malaria from microscopic red

blood cell images☆
Mosabbir Bhuiyan a, Md Saiful Islam b, *
a
Department of Electronics and Telecommunication Engineering, Chittagong University of Engineering and Technology, Chattogram, 4349, Bangladesh
b
Department of Electrical and Electronic Engineering, Chittagong University of Engineering and Technology, Chattogram, 4349, Bangladesh

A R T I C L E I N F O A B S T R A C T

Keywords: Malaria is a life-threatening parasitic disease spread by infected female Anopheles mosquitoes. After analyzing it,
Machine learning microscopists detect this disease from the sample of microscopic red blood cell images. A professional micros-
Deep learning copist is required to conduct the detection process, such an analysis may be time-consuming and provide low-
CNN
quality results for large-scale diagnoses. This paper develops an ensemble learning-based deep learning model
Transfer learning
Ensemble
to identify malaria parasites from red blood cell images. VGG16(Retrained), VGG19(Retrained), and DenseNe-
Malaria t201(Retrained) are three models that are used in developing the adaptive weighted average ensemble models. To
reduce the dispersion of predictions, a max voting ensemble technique is then applied in combination with
adaptive weighted average ensemble models. A variety of image processing techniques are utilized including the
data augmentation technique to increase the number of data and solve the overfitting problem of the model. Some
other approaches of custom CNN, Transfer Learning, and CNN-Machine Learning (ML) classifier techniques are
also implemented for comparing their performance with the ensemble learning model. The proposed ensemble
learning model provides the best performance among all with an accuracy of 97.92% to classify parasitized and
uninfected cells. Therefore, the deep learning model has the potential to diagnose malaria more accurately and
automatically.

1. Introduction Malaria killed 627,000 individuals in 2020. Children under age 5 are in
the most dangerous position. In 2020, 80% of children worldwide died
Malaria is a serious disease that can be fatal. An infected Anopheles from malaria [20]. Plasmodium parasites are the cause of malaria.
mosquito bite is the most common way to contract the disease. Infected Humans can be infected by five different malaria parasites. These are
mosquitos carry plasmodium parasites. Malaria is transferred when a Plasmodium falciparum, P. vivax, P. ovale, P. malariae, and P. knowlesi.
mosquito bites an infected person and then bites a non-affected person. Malaria signs typically appear 10 to 4 weeks after the infection has
When parasites get mature, they travel to the liver and enter the blood- occurred. It can take several months for symptoms to appear in some
stream, where they begin to infect red blood cells after a few days. Ma- cases. Some common signs of malaria disease include moderate to severe
laria can also be transmitted through organ transplants, blood shivering chills, high fever, muscle pain, profuse sweating, headache,
transfusions, and the use of blood-contaminated syringes and needles. vomiting, bloody feces, abdominal pain, diarrhea, nausea, and convul-
Malaria seems to be the most common disease in tropical and subtropical sions [5]. Because of the loss of red blood cells, malaria can induce
regions, where the parasites proliferate. According to the most recent jaundice (yellow skin and eyes) and anemia. If not treated promptly, the
World Malaria Report, estimated 241 million malaria cases in 2020, infection can cause kidney failure, mental confusion, coma, convulsions,
slightly higher compared to 227 million in 2019. It was estimated that and death [6]. To ensure whether anyone has malaria disease or not,

☆
Here, R denotes that Transfer Learning models are retrained.
* Corresponding author.
E-mail address: [email protected] (M.S. Islam).

Production and hosting by Elsevier

https://doi.org/10.1016/j.sintl.2022.100209
Received 9 August 2022; Received in revised form 4 November 2022; Accepted 8 November 2022
Available online 21 November 2022
2666-3511/© 2022 The Authors. Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 3. Resized image of 64x64 pixels.

Table 1
Dataset splitting into training, validation, and testing
sets.
Splitting Type Percentage

Training 70%
Validation 10%
Testing 20%

Therefore, there are several benefits of using an automated malaria

detection process. It provides much higher accurate results compared to
the manual process. It reduces the workload to serve more patients. Deep
learning has that much potential to provide an accurate and faster result
for parasites.
As malaria is a life-threatening disease, many researchers have been
working on this for a couple of years. Previously, malaria is detected by
human expertise. Then it was tough for an inexperienced human to detect
the disease correctly. But now researchers have got a tremendous result
on this. Computer-aided diagnosis is the most powerful tool in this area.
A variety of image processing techniques are performed to extract fea-
Fig. 1. Proposed methodology. tures from images and a classifier is then used to identify disease. In
Ref. [33], the author proposed a conventional image processing method.
Red blood cells are segmented using the Otsu binarization technique. He
there are several tests available. The most common and accurate test is
extracted statistical features manually from the images. To cluster image
the thick and thin blood smear test. Microscopists determine the parasites
values, K-means clustering is applied. Support vector machine (SVM) is
from red blood cells using a microscope. This test must require trained
used as a classifier to classify the disease. In Ref. [24], the author showed
and experienced microscopists. Another is a Rapid diagnostic test or
a variety of approaches for malaria diagnosis. Image processing, cell
antigen test. When microscopy is not available, a rapid diagnostic test can
segmentation, feature extraction, and classification can all be done by a
be performed instead of blood smears. This test identifies malaria anti-
variety of techniques. Preprocessing methods he mentioned mean filter,
gens from a person’s blood and changing the color of the testing result
Laplacian filter, wiener filter, median filter, adaptive histogram equal-
gives a positive result. It also has no way of knowing whether the
ization, contrast enhancement, etc. Segmentation techniques are
infection is minor or severe. Another is the Molecular test or Polymerase
morphological operation, Hough transformation, k-means clustering,
chain reaction (PCR). The PCR is a laboratory tool to detect and identify
Fuzzy segmentation, etc. Features that are mentioned in his paper are a
Plasmodium species from parasite DNA. In laboratories where there is a
color feature, texture feature, and morphologic feature. In the classifi-
limitation of training and expertise in microscopic examination of ma-
cation method, he included supervised and unsupervised learning algo-
laria, a PCR test might be a preferable option to begin the treatment. The
rithms. Deep learning is the most prominent way to detect the disease
two other tests are the Antibody test and the Susceptibility test [4].

Fig. 2. Microscopic red blood cell images of Malaria.

2
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

coefficient). He found an accuracy of 95.32% in his research. In Ref. [7],

the author proposed a three-stage pipeline that is segmentation, crop &
musk, and classification. In the segmentation stage, a segmentation
neural network (SNN) [2] has been built for red blood cell (RBC) seg-
mentation. Another 13 layered CNN architecture has been created to
classify malaria disease. He got an overall 93.72% accuracy from his
custom CNN architecture. In Ref. [9], the author introduced a new
approach to classify malaria disease that is CNN-SVM and CNN–KNN. In
this approach, CNN is employed as a feature extractor and SVM/KNN is
applied as a classifier. In addition, he demonstrated a new autoencoder
technique. Autoencoder is a form of neural network that is trained to
regenerate output from its input. Again in Ref. [8], the author proposed a
pre-trained network technique. He employed LeNet [17,18], AlexNet
[15], and GoogLeNet [34] in his research. A different dataset was
collected from the pathologists of the University of Alabama at Bir-
mingham. Pre-trained networks were used as a feature extractor. Using
KUllback-Leibler (KL) distance, the best seven features were chosen from
76 features. He applied an SVM classifier to classify malaria. In Ref. [21],
the author proposed a custom CNN architecture. Bilateral filter & data
augmentation techniques were applied as data preprocessing steps. The
preprocessed input images were then fed into a custom-built CNN model.
He found an accuracy of 96.82%. In Ref. [25], the author introduced a
unique CNN architecture called Attentive Dense Circular Net (ADCN) in
their research. ADCN was inspired by ResNet [11] and DenseNet [12].
There are three parts to the ADCN architecture. In the first part, there is a
custom CNN model with two dense blocks and one attention module. The
dense block has six dense conv block that consists of two convolutional
layers. Every dense conv block is connected with another dense conv
block. Finally, an attention module is split into two parts. One is a 1x1
convolutional layer, while the other is an attention branch for extracting
the attention feature map. To ensure that the branch's output size is
consistent, the attention branch uses downsampling and upsampling
techniques. In Ref. [26], the author proposed three different techniques
that are custom CNN network, transfer learning as VGG16 [31], and
CNN-SVM. Then he used the ensemble learning method to reduce pre-
diction variance and generalization error. The results from each of the
three networks are combined, and the final prediction is calculated using
a weighted average ensemble. In Ref. [1], the author proposed a novel
Fig. 4. Augmented cell images of Malaria.
method known as Incremental Modular Networks (IMNets) to classify
malaria. This method combines multiple SubNet to provide additional
information. Each SubNets module is added iteratively, either serially or
Table 2
parallelly, to the current architecture. In Ref. [10], the author discussed
Augmented types and their values.
some protein sequence formulation strategies such as discrete methods,
Augmented type Value
biochemical, physiochemical, and natural language processing tech-
Rotation 0.2 niques. These are used to convert protein sequences into numerical de-
Zoom 0.2 scriptors. Four classification methods are applied, and the predicted
Horizontal flip True
outcomes of these classifiers are then merged to develop an ensemble
Vertical flip True
Rescale 1/255 model using majority and genetic algorithm.

2. Methodology
automatically and more correctly. Nowadays, the most often used
approach is the convolutional neural network (CNN) [16]. It has the An ensemble learning-based system is developed in this research to
potential to extract features from images automatically. In Ref. [22], the reduce the burden on microscopists. Ensemble learning is a technique
author proposed a fast Convolutional Neural Network of six convolu- where multiple models are trained on the same dataset and combined
tional layers, one fully connected layer, and one classification layer. He with their results to achieve better accuracy and reduce the variance of
also evaluated different transfer learning techniques such as AlexNet the model. The input data is initially resized into 64 pixels of height and
[15], VGG16 [31], ResNet50 [11], DenseNet121 [12]. Finally, he pro- 64 pixels of width. Then, the dataset is divided into 70% training, 10%
vided an overall accuracy of 96.7% for all models using the test dataset. validation, and 20% testing sets. After that data augmentation technique
Transfer learning is a way of using pre-trained networks. He found an is applied with a variety of domains to increase the number of training
accuracy of 95.9% using the transfer learning model. In Ref. [29], the data which enhances the model’s performance. VGG16(R), VGG19(R),
author developed a shallow CNN model that performs similarly to the and DenseNet201(R) models are fine-tuned using several hyper-
VGG16 [31] and ResNet50 [11]. In deep learning, high computational parameters such as optimizer, learning rate, activation function, etc.*
complexity and high computational cost is a big issue. His proposed Then, these models are trained to learn patterns in the data using a
custom CNN has lower computational complexity and lower computa- training set. From the individual model, two of the models are taken and
tional run time. Evaluation metrics used in his research are accuracy, applied adaptive weighted average ensemble to achieve better perfor-
sensitivity, specificity, F1 score, and MCC (Mathew’s Correlation mance by reducing the variance of the model. This method finds the best

3
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 5. Transfer Learning model.

cells. The image resolutions are varying between 110 and 150 pixels. In
Table 3 the preprocessing section, the images are scaled into a fixed range of 64
Hyperparameter and their values in Transfer Learning. pixels. Researchers at the Lister Hill National Center for Biomedical
Parameter Name Type/Value Communications (LHNCBC), which is a part of the National Library of
Input Size 64 x 64 Medicine (NLM), have built an Android mobile application attached to a
Batch Size 32 conventional light microscope. Giemsa-stained thin blood smear slides
Epochs 100 were provided by 150 Plasmodium falciparum-infected patients and 50
Pre-trained weights ImageNet healthy people. Then, images were captured from those samples by using
Activation Function Softmax
Optimizer Adam
a smartphone at Chittagong Medical College Hospital, Bangladesh. A
Learning Rate 1e-04 professional slide reader annotated the images manually at the
Loss Function Sparse Categorical Cross entropy Mahidol-Oxford Tropical Medicine Research Unit in Bangkok, Thailand
(NIH Malaria Dataset, LHNCBC). Fig. 2 shows the parasitized and unin-
fected microscopic red blood cell images of malaria.
outcome with the appropriate weights automatically. VGG16(R) and
According to Ref. [9], there are some mislabeled data in the dataset.
VGG19(R) achieve a higher result with 0.7 and 0.6 wt respectively. With
Some of the data are labeled as parasitized but they are uninfected. In
0.6 and 0.5 wt, VGG19(R) and DenseNet201(R) produce a better result.
addition, some of the data from uninfected cells are labeled as parasit-
When compared to other weights, VGG16(R) and DenseNet201(R)
ized. The expert also verified that some of the data is incorrectly labeled.
perform better with 0.4 and 0.5 wt, respectively. Three ensembled
There are found 647 falsely labeled parasitized data and 750 falsely
models are obtained after performing an adaptive weighted average
uninfected data [9]. Around 5% of data is mislabeled in the total dataset
technique. After that, another ensemble technique known as max voting
which may occur an impact on the performance of the model.
is applied using three ensembled models with hard voting to improve the
model's performance and reduce the dispersion of the predictions. The
2.2. Environmental setup
whole technique of this study is depicted in Fig. 1.
All the experiments are performed in the Google Colab platform with
13 GB Random Access Memory (RAM), 69 GB Colab Disk, Keras® 2.4.3
2.1. Dataset
API with Tensorflow® 2.5.0 backend, Python® 3.7.10, Nvidia® Tesla T4
Graphical Processing Unit (GPU), CUDA V11.2 dependencies for GPU
The dataset is collected from the National Institutes of Health (NIH)
acceleration and matplotlib® 3.2.2 library.
for this research [23]; LHNCBC). There are 27,558 cell images in total. It
is classified into two sections: 13,779 parasitized and 13,779 uninfected

Fig. 6. Ensemble of ensemble method.

4
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 7. Custom CNN architecture.

Table 4
Layer's name, shape, and parameter values in Custom CNN.
Layer Output Shape Parameter

Conv2D 64 x 64 x 32 2432
ReLu 64 x 64 x 32 0
Max pooling 32 x 32 x 32 0
Batch Normalization 32 x 32 x 32 128
Conv2D 32 x 32 x 32 9248
ReLu 32 x 32 x 32 0
Max pooling 16 x 16 x 32 0
Batch Normalization 16 x 16 x 32 128
Conv2D 16 x 16 x 64 18496
ReLu 16 x 16 x 64 0
Max pooling 8 x 8 x 64 0
Batch Normalization 8 x 8 x 64 256
Flatten 4096 0
Dense 1024 4195328
Dropout 1024 0
Dense 2 2050
Total Parameter 4228066

Table 5
Hyperparameter and their values in Custom CNN.
Parameter Name Type/Value

Input Size 64 x 64
Batch Size 32
Epochs 100
Regularization 0.001
Dropout 0.2
Activation Function Softmax
Optimizer Adam
Learning Rate 1e-04
Loss Function Sparse Categorical Cross entropy

Fig. 9. Custom CNN model performance (a) with augmentation and (b) without
augmentation.

2.3. Data preprocessing & splitting

The sizes of malaria images are varying from 110 pixels to 150 pixels.
In this experiment, the cell images are scaled to 64 pixels in height and 64
pixels in width. Fig. 3 includes the resized cell images of malaria. Dataset
Fig. 8. CNN-ML classifier model. splitting is the general part of model building. The amount of the data
impacts how the dataset is split [3]. In this research, the dataset is split

5
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 10. Custom CNN model performance at different optimizers.

counterclockwise. The value of 0.2 is used for the zoom transformation

Table 6 which denotes random zooming vertically with upper and lower bound
Result performance of custom CNN model at different optimizers. in the range of 20% [28]. The resulting image is represented in the form
Optimizer Precision Recall F1 score Accuracy of zooming in and zooming out transformations. Horizontal and vertical
Adadelta 0.7794 0.7632 0.7606 0.7643
flips are used to shift images from left to right and top to bottom. Images
Adagrad 0.9122 0.9122 0.9121 0.9121 are rescaled to map features in the range of 0 to 1 by passing scale ¼
Adamax 0.9616 0.9603 0.9604 0.9605 1/255 to get faster convergence.
Ftrl 0.2476 0.5000 0.3312 0.4953
Nadam 0.9700 0.9692 0.9693 0.9694
RMSprop 0.9634 0.9634 0.9634 0.9634 2.5. Transfer learning models
SGD 0.9456 0.9422 0.9424 0.9425
Adam 0.9723 0.9718 0.9719 0.9719 Transfer learning models such as VGG16(R), VGG19(R), and Dense-
Net201(R) are employed to apply the adaptive weighted average
ensemble in the proposed method. Here models are retrained so that the
into 70% of the training set, 10% of the validation set, and 20% of the
weights are updated and assist the network in learning and identifying
testing set using scikit learn library. The model is trained to find patterns
features associated with new images and categories. In this approach, the
in the data by utilizing the training set. The validation set is used to test
final dense layer is eliminated as transfer learning models are created for
the model’s performance during training. Finally, the model is tested
1000 categories. A new output dense layer is added with 2 neurons to
with a test set when it has completed training. The chart diagram of
classify parasitized and uninfected cells. To achieve the best outcomes,
dataset splitting is shown in Table 1.
hyperparameters are used to fine-tune the transfer learning models.
Adam optimizer is one of the hyperparameters which is used with a
2.4. Data augmentation learning rate of 1e-04 to reduce the error by updating weights and biases.
As a loss function, sparse categorical cross-entropy is applied to calculate
Data augmentation is the process of creating new artificial data from the error between the ground and predicted values. The batch size and
the existing data. It avoids unwanted features from being learned and the epochs number are 32 and 100 respectively. The architecture is given
improves overall performance. Deep learning requires a larger dataset for with all of the transfer learning models in Fig. 5 and their hyper-
training so that the model can learn more patterns from images and make parameters value is listed in Table 3.
accurate predictions. The data augmentation technique increases the
quantity of the training dataset with different transformations and en- 2.6. Ensemble of ensemble method
hances the performance of the model. Transforms of data include some
operations such as rotating, zooming, flipping, rescaling, shearing, etc. In In this step, an ensemble of ensemble method is applied to get the best
this research, the data augmentation technique is applied to overcome performance. After taking three individual transfer learning models, two
the overfitting problem where a model learns effectively in the training models are combined every time to apply the adaptive weighted average
set but not in the test set. Fig. 4 shows the transformation of data method. As a result, three different adaptive weighted average models
augmentation techniques that are applied to the training dataset and are discovered. Weighted average model 1 is formed by combining
Table 2 shows their respected values. Random rotation is applied with VGG16(R) and VGG19(R), while weighted average model 2 is developed
the value of 0.2 which describes a result of random rotation with upper by combining VGG19(R) and DenseNet201(R). Weighted average model
and lower bound in the range of [
20% * 2pi, 20% * 2pi] [27]. The 3 is created by using VGG16(R) and DenseNet201(R). In a normal
resulting image is rotating randomly both clockwise and weighted average method, weights need to be initialized to get the best

6
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 11. Custom CNN model performance at different Learning rates.

1-pixel stride to maintain the same dimension of the feature map. The
Table 7 output of a convolution layer is processed through a ReLu activation
Result performance of custom CNN model at different Learning rates. function to introduce non-linearity. It speeds up the training process and
Learning Rate Precision Recall F1 score Accuracy converges faster [30]. It also allows neurons to learn arbitrary complex
1e-02 0.2476 0.5000 0.3312 0.4953
transformations. The ReLu activation function is shown in Equation (1),
1e-03 0.9550 0.9547 0.9545 0.9545 where negative values become zero and positive values remain the same.
1e-04 0.9723 0.9718 0.9719 0.9719
RðxÞ ¼ max ð0; xÞ (1)
Here, R is the result after adding non-linearity to matrix x.
Table 8 Then a max-pooling layer is applied with a 3x3 matrix and a stride of
Result performance of different CNN-ML classifiers. 2 to reduce the dimension of images from the output of a convolutional
layer. It also reduces the computational cost by minimizing the number of
Model Precision Recall F1 score Accuracy
parameters. After that, a batch normalization layer is used to convert
CNN-SVM 0.8171 0.8165 0.8166 0.8167
numerical data to a common scale without changing the shape of the
CNN–KNN 0.6458 0.6155 0.5958 0.6177
CNN-Decision Tree 0.7386 0.7386 0.7386 0.7386
data. It solves the overfitting problem and makes neural network faster as
CNN-Random Forest 0.8132 0.8123 0.8119 0.8120 well as more stable.

1 X ðiÞ
μ¼ N (2)
result. However adaptive weighted average method automatically n i
chooses appropriate weights based on the data. In weighted average
model 1, VGG16(R) and VGG19(R) produce a better outcome with 0.7 1 X ðiÞ 
and 0.6 wt respectively. With 0.6 and 0.5 wt, VGG19(R) and Dense- σ¼ N
μ (3)
n i
Net201(R) achieve a higher result in weighted average model 2. In
weighted average model 3, VGG16(R) and DenseNet201(R) perform
ðiÞ N ðiÞ
μ
better with 0.4 and 0.5 wt, respectively. Finally, the max voting ensemble Nnorm ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi (4)
method is applied to the three weighted average models where the ma- σ2
ε
jority class is the final output. The ensemble of ensemble method is
ðiÞ
shown in Fig. 6. N ¼ γ*Nnorm þβ (5)

Using equations (2) and (3), the Batch normalization layer calculates
2.7. Other architectures the mean μ and the standard deviation σ of the activation values
throughout the batch. The activation vector N ðiÞ is then normalized with
2.7.1. Custom CNN architecture equation (4). Finally, it computes the output of the layer N ðiÞ by applying
The custom CNN architecture is developed with a total of 16 layers a linear transformation with two trainable parameters γ and β using
which has 3 blocks of convolutional, activation, max pooling, and batch equation (5) [13]. The second and third convolutional layer consists of
normalization layers, 1 flatten, and dropout layer, and the last 2 dense or 32 and 64 kernels of a 3x3 matrix respectively. The ReLu activation
fully connected layers. The first convolutional layer performs convolved function is used to make all linear convolutional output non-linear. There
operations with 32 kernels to extract feature maps from the input images. is a max-pooling and batch normalization layer with the same value after
The kernel is applied with a 5x5 matrix including the same padding and

7
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Fig. 12. Transfer learning models performance.

Fig. 13. Retrained transfer learning models performance.

each convolutional layer. neurons with a ReLu activation function to transform linear input to non-
The feature extraction part is completed in the upper section and now linear. L2 regularization is applied in that layer to minimize the value of
this section will describe classifying malaria parasites by learning pa- weighted metrics. Therefore, it solves the overfitting problem by
rameters. Previously, the data is stored as a three-dimensional feature reducing the weighted values. These values update the cost function
vector. It must be converted into a one-dimensional array to classify which is shown in equation (6).
parasitized and uninfected cells. So, flatten layer is used to make the
three-dimensional array into a one-dimensional linear array to feed data λ X 2
Cost function ¼ Loss þ * jjwjj (6)
in a fully connected layer. It consists of 4096 neurons and this is the first 2m
fully connected layer or dense layer of the model. The output from flatten Here, λ is the regularization parameter.
layer is passed to the next dense layer. The hidden dense layer has 1024 Dropout is another regularization technique that is also included with

8
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

Table 9 Table 11
Result performance of different transfer learning models. Result comparison between proposed ensemble model and other models.
Model Precision Recall F1 score Accuracy Model Precision Recall F1 score Accuracy

VGG16 0.9206 0.9179 0.9181 0.9182 Custom CNN 0.9723 0.9718 0.9719 0.9719
VGG19 0.9124 0.9101 0.9103 0.9104 VGG16 (R) 0.9752 0.9752 0.9752 0.9752
DenseNet201 0.9066 0.8920 0.8918 0.8929 CNN-SVM 0.8171 0.8165 0.8166 0.8167
ResNet50V2 0.9034 0.9001 0.9002 0.9005 Ensemble 0.9791 0.9792 0.9792 0.9792
VGG16 (R) 0.9752 0.9752 0.9752 0.9752
VGG19 (R) 0.9739 0.9739 0.9739 0.9739
DenseNet201 (R) 0.9740 0.9739 0.9739 0.9739 Table 12
ResNet50V2 (R) 0.9707 0.9706 0.9706 0.9706 Result Comparison between proposed ensemble model and existing models.
Model Dataset Precision Recall F1 Accuracy
score

Otsu segmentation, K- NIH 0.9607 0.93 0.945 0.946

means clustering [33]
CNN [22] NIH – – – 0.96
VGG16 [29] NIH – – – 0.9615
CNN [21] NIH 0.9682 0.9633 0.9682 0.9682
VGG16 [26] NIH 0.9719 0.9720 0.9709 0.9777
VGG16 [35] NIH – 0.956 0.956 0.96
CNN [19] NIH 0.9773 0.9699 0.9736 0.9737
Proposed Ensemble NIH 0.9791 0.9792 0.9792 0.9792
model

truth values. The model is trained using 32 training samples and then run
for 100 epochs. To minimize the error, hyperparameters are fine-tuned
using their values, which are listed in Table 5.

2.7.2. CNN-ML classifier

In this section, the custom CNN architecture is used as a feature
extractor and several machine learning algorithms are employed as a
classifier to compare their performance with the softmax classifier [32].
Some of the machine learning algorithms such as SVM, KNN, Decision
Tree, and Random Forest are applied in this research. SVM is a supervised
Fig. 14. Confusion matrix of the proposed ensemble method. learning algorithm, that gives effective results when the number of di-
mensions exceeds the number of samples. Radial Basis Function (RBF) is
applied in SVM classifier as a kernel whose value depends on the distance
a value of 0.2 in the next dense layer to deactivate 20% of nodes
from the origin. Equation (8) shows RBF kernel mathematical expression.
randomly. It also helps to avoid the problem of overfitting. Lastly, a
classification dense layer is added with 2 neurons to classify parasitized  
KðX1 ; X2 Þ ¼ exp
γjjX1
X2 jj2 (8)
and uninfected cells from microscopic red blood cell images. A softmax
activation function is applied in that layer as a classifier to give results as
Where, X1 and X2 are the variables and jjX1
X2 jj is Euclidean distance
a probability distribution. The expression of the softmax activation
between X1 & X2 .
function is given in equation (7).
A Random Forest classifier that combines multiple decision trees
exp ðzi Þ because of that 100 trees are applied to solve a complex problem. KNN
Sðzi Þ ¼ P   (7) and Decision Tree are also used to evaluate the model's performance.
j exp zj
Fig. 8 represents the CNN-ML classifier diagram.

S is the softmax probability result and z represents the values of 3. Result analysis
output layer neurons.
In this research, different deep learning methods are tested to find a
Fig. 7 shows the custom CNN architecture and Table 4 represents the more accurate and efficient model for identifying malaria from red cell
layer's name, output, and parameter values. images. The results are evaluated based on different optimizers, learning
The model is fine-tuned using the Adam optimizer, which has a rates, transfer learning models, ML classifiers, and ensemble learning
learning rate of 1e-04 and updates weights and biases to minimize losses. methods. All of the experiments are run with a learning rate of 1e-04, a
Sparse categorical cross-entropy is utilized as a loss function for binary batch size of 32. Precision, recall, F1 score, and accuracy are utilized as
classification. It determines the error between the predicted and ground performance metrics to evaluate the model's performance. There are four

Table 10
Result performance of different ensemble learning methods.
Ensemble method Model Weights Precision Recall F1 score Accuracy

Adaptive Weighted Average VGG16(R), VGG19(R) 0.7, 0.6 0.9759 0.9759 0.9759 0.9759
VGG19(R), DenseNet201(R) 0.6, 0.5 0.9784 0.9784 0.9784 0.9784
VGG16(R), DenseNet201(R) 0.4, 0.5 0.9781 0.9781 0.9781 0.9781
VGG16(R), VGG19(R) & DenseNet201(R) 0.1, 0.9, 0.8 0.9786 0.9786 0.9786 0.9786
Max Voting VGG16(R), VGG19(R) & DenseNet201(R) – 0.9773 0.9773 0.9773 0.9773
Adaptive Weighted Average & Max Voting VGG16(R), VGG19(R) & DenseNet201(R) 0.9791 0.9792 0.9792 0.9792

9
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

important parameters. They are true positive (TP), false positive (FP), 3.1.4. Ensemble learning performance
true negative (TN), and false negative (FN). True positive denotes Fig. 14 represents the confusion matrix of parasitized and uninfected
infected cells classified correctly as infected. False positive defines classes for the proposed ensemble method. The proposed ensemble
infected cells classified as uninfected. True negative describes uninfected method is a combination of adaptive weighted average and max voting
cells correctly classified and false negative is the opposite of this. In ensemble techniques. This method correctly classified 2678 images from
equations (9)–(12), various evaluation metrics are shown. the parasitized class and 2723 images from the uninfected class. The
testing accuracy is found at 97.92% for this ensemble method.
TP Table 10 shows the performance results of different ensemble
Precision ¼ (9)
TP þ FP learning methods. The adaptive weighted average approach gives the
best result for a given set of weights automatically. The proposed
TP
Recall ¼ (10) methodology is the combination of two ensemble methods that is
TP þ FN adaptive weighted average and max voting. This ensemble technique
provides the best result among all strategies.
2  Precision  Recall
F1 Score ¼ (11)
Precision þ Recall
3.2. Comparison
TP þ TN
Accuracy ¼ (12) 3.2.1. Result comparison with different models
TP þ FP þ TN þ FN
The outcomes are now compared to other deep learning models,
including the proposed ensemble learning method. The ensemble model
3.1. Performance
is the best model among others and achieves a high testing accuracy of
97.92%. Table 11 shows the performance comparison between the pro-
3.1.1. Custom CNN performance
posed ensemble model and other models using several evaluation
The custom CNN model achieves 96.43% of training accuracy and
metrics.
96.56% of validation accuracy. There is no overfitting problem as the
data augmentation technique is applied. Without data augmentation, the
3.2.2. Result comparison with existing methods
training accuracy is found at 99.77% and the validation accuracy is
In this section, the result is compared with existing works based on
95.06%. As can be seen, the training and validation accuracy results are
different performance metrics. Many researchers have proposed the
vastly different. This is called the overfitting problem where the model
transfer learning model as well as the CNN model in their papers.
performs well in training data but not in test data. Fig. 9 shows the
However, the results are not as good as the proposed method. The pro-
performance of the custom CNN model with and without augmentation.
posed ensemble model obtains a significant testing accuracy of 97.92%
There are a series number of optimizers available such as Adadelta,
and detects malaria more accurately and efficiently. The performance
Adagrad, Ftrl, SGD, Adam, RMSprop, etc. Fig. 10 shows the performance
comparison between the proposed ensemble method and existing
of those optimizers in the custom CNN model. Adam outperforms other
methods is shown in Table 12.
optimizers because it is computationally efficient and uses less memory.
It is invariant to diagonal rescaling of gradients and is ideally suited to
4. Discussion
large data/parameter problems [14]. Table 6 represents the performance
results of the custom CNN model at different optimizers.
This study introduces a novel method of detecting malaria parasites
Fig. 11 demonstrates the performance of the custom CNN model at
by combining two ensemble learning methods to create a model. Previ-
different learning rates. An optimal learning rate of 1e-04 is required to
ously, there are so many models developed by many researchers. In their
reach in global minima point. At this point, the loss is minimum, and
research, they built either a custom CNN model or a transfer learning
identify malaria parasites more accurately. Because the two other
model. They also attempted to construct an ensemble learning model. But
learning rate values are too high, they are unable to achieve global
in their model, they only applied one ensemble technique to identify
minima point. Table 7 shows the performance results of the custom CNN
malaria and the performance of those models is not very accurate.
model at different learning rates.
However, in this study, two ensemble techniques are combined to create
a novel architecture. First, transfer learning models are used to apply
3.1.2. CNN-ML classifier performance
weighted average ensemble, and then the max voting ensemble tech-
The outcomes of various machine learning algorithms are now being
nique is applied to weighted average models to provide more accurate
examined. Custom CNN is used as a feature extractor and machine
results. When compared to previous work, this unique model performs
learning algorithms are applied as a classifier. CNN-SVM is performing
better and more accurately, as seen in Table 12.
better than the three other CNN-ML classifiers. The test accuracy is found
81.67% for the SVM classifier. As compared to the softmax classifier,
5. Conclusion and future work
these ML classifiers are not performing very well to classify malaria
parasites. Table 8 shows the performance results for different CNN-ML
In this paper, an ensemble learning-based deep neural network is
classifiers.
proposed to classify malaria parasites from microscopic red blood cell
images. In addition, several deep learning methods are evaluated
3.1.3. Transfer learning performance
including the custom CNN model, transfer learning model, and CNN-ML
Fig. 12 illustrates the performance of the transfer learning models that
classifier model to compare their performance with the proposed
use pre-trained weights, while Fig. 13 shows the performance of the
ensemble method. The proposed ensemble learning model provides the
models that have been retrained and whose weights have been updated.
best performing result in classifying malaria parasites with a testing ac-
From the figures, models that use pre-trained weights do not perform
curacy of 97.92%. It also minimizes the dispersion of predictions and
very well. However, models that have been retrained, perform incredibly
improves model performance. The National Institutes of Health (NIH)
well since all layers are trained and the networks learn new attributes
provided the dataset for this study. A variety of preprocessing techniques
associated with new images and categories. VGG19 (R) outperforms the
are being performed to improve the model's performance. Data
others by achieving a testing accuracy of 97.52%. Table 9 shows the
augmentation technique is one of them which is applied to solve the
performance results for different transfer learning models.
overfitting problem. The model is fine-tuned using various hyper-
parameters to get the minimum error in classification. Therefore, the

10
M. Bhuiyan, M.S. Islam Sensors International 4 (2023) 100209

ensemble model performs exceptionally well and significantly acceler- [15] A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep
convolutional neural networks, Adv. Neural Inf. Process. Syst. 25 (2012)
ates the detection process. The proposed methodology will assist mi-
1097–1105. http://code.google.com/p/cuda-convnet/.
croscopists in the field of biomedical in detecting malaria disease more [16] Y. Lecun, Y. Bengio, Convolutional networks for images, speech, and time-series,
correctly and effectively. In the future, a new architecture will be Handb.Brain.Theor. Neural Network. 3361 (10) (1995) 255–258. https://dl.acm.
developed to increase detection performance. The evaluation of various org/doi/10.5555/303568.303704.
[17] Y. Lecun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to
architectures on a variety of malaria datasets will continue. document recognition, Proc. IEEE 86 (11) (1998) 2278–2324.
[18] Y. LeCun, K. Kavukcuoglu, C. Farabet, Convolutional networks and applications in
Declaration of competing interest vision, in: ISCAS 2010 - 2010 IEEE International Symposium on Circuits and
Systems: Nano-Bio Circuit Fabrics and Systems, 2010, pp. 253–256, https://
doi.org/10.1109/ISCAS.2010.5537907.
The authors declare that they have no known competing financial [19] Z. Liang, A. Powell, I. Ersoy, M. Poostchi, K. Silamut, K. Palaniappan, P. Guo,
interests or personal relationships that could have appeared to influence M.A. Hossain, A. Sameer, R.J. Maude, J.X. Huang, S. Jaeger, G. Thoma, CNN-based
image analysis for malaria diagnosis, in: In 2016 IEEE International Conference on
the work reported in this paper. Bioinformatics and Biomedicine (BIBM), 2016, pp. 493–496.
[20] Malaria, World Health organization (WHO), Retrieved from, https://www.who.int/
References news-room/fact-sheets/detail/malaria, 2021. (Accessed 25 January 2022).
[21] A. Maqsood, M.S. Farid, M.H. Khan, M. Grzegorzek, Deep malaria parasite detection
in thin blood smear microscopic images, Appl. Sci. 11 (5) (2021) 2284, https://
[1] R. Ali, R.C. Hardie, B.N. Narayanan, T.M. Kebede, IMNets: deep learning using an
doi.org/10.3390/app11052284.
incremental modular network synthesis approach for medical imaging applications,
[22] B.N. Narayanan, R. Ali, R.C. Hardie, Performance analysis of machine learning and
Appl. Sci. 12 (11) (2022), https://doi.org/10.3390/app12115500.
deep learning architectures for malaria detection on cell images, in: Applications of
[2] V. Badrinarayanan, A. Kendall, R. Cipolla, SegNet: a deep convolutional encoder-
Machine Learning, 2019, September 6, https://doi.org/10.1117/12.2524681.
decoder architecture for image segmentation, IEEE Trans. Pattern Anal. Mach.
[23] NIH malaria dataset, LHNCBC, Retrieved from, https://lhncbc.nlm.nih
Intell. 39 (12) (2017) 2481–2495, https://doi.org/10.1109/TPAMI.2016.2644615.
.gov/LHC-downloads/downloads.html#malaria-datasets. (Accessed 27 January
[3] M. Bhuiyan, A. Kabir, M. Ahasan Kabir, Vehicle speed prediction based on road
2022).
status using machine learning, Adv. Res.Eng. Energy 2 (1) (2020) 1–9. https://fazp
[24] M. Poostchi, K. Silamut, R.J. Maude, S. Jaeger, G. Thoma, Image analysis and
ublishing.com/aree/index.php/aree/article/view/11.
machine learning for detecting malaria, Transl. Res. 194 (2018) 36–55, https://
[4] Blood Smear - Understand the Test, Retrieved from, https://www.testing.com/test
doi.org/10.1016/j.trsl.2017.12.004. Mosby Inc.
s/blood-smear, 2021. (Accessed 26 January 2022).
[25] Q. Quan, J. Wang, L. Liu, An effective convolutional neural network for classifying
[5] D. Burke, Malaria: causes, symptoms, and diagnosis. Healthline, Retrieved from,
red blood cells in malaria diseases, Interdiscipl. Sci. Comput. Life Sci. 12 (2) (2020)
https://www.healthline.com/health/malaria#symptoms, 2019. (Accessed 25
217–225, https://doi.org/10.1007/s12539-020-00367-7.
January 2022).
[26] A. Rahman, H. Zunair, M. Sohel Rahman, J. Quader Yuki, S. Biswas, M. Ashraful
[6] CDC - Malaria - About Malaria - FAQs, Global Health, Division of Parasitic Diseases
Alam, N. Binte Alam, M.R.C. Mahdy, Improving Malaria Parasite Detection from
and Malaria, 2021. https://www.cdc.gov/malaria/about/faqs.html.
Red Blood Cell Using Deep Convolutional Neural Networks, 2019 arXiv Preprint
[7] M. Delgado-Ortet, A. Molina, S. Alferez, J. Rodellar, A. Merino, A deep learning
arXiv:1907.10418.
approach for segmentation of red blood cell images and malaria detection, Entropy
[27] RandomRotation layer, Keras, Retrieved from, https://keras.io/api/layers/preproce
22 (6) (2020) 657, https://doi.org/10.3390/e22060657.
ssing_layers/image_augmentation/random_rotation. (Accessed 27 January 2022).
[8] Y. Dong, Z. Jiang, H. Shen, W. David Pan, L.A. Williams, V.V.B. Reddy,
[28] RandomZoom layer, Keras, Retrieved from, https://keras.io/api/layers/preproce
W.H. Benjamin, A.W. Bryan, Evaluations of deep convolutional neural networks for
ssing_layers/image_augmentation/random_zoom. (Accessed 27 January 2022).
automatic identification of malaria infected cells, in: In 2017 IEEE EMBS
[29] S. Sarkar, R. Sharma, K. Shah, Malaria Detection from RBC Images Using Shallow
International Conference on Biomedical & Health Informatics (BHI), 2017,
Convolutional Neural Networks, 2020 arXiv Preprint arXiv:2010.11521.
pp. 101–104, https://doi.org/10.1109/BHI.2017.7897215.
[30] W. Shang, K. Sohn, D. Almeida, H. Lee, Understanding and improving convolutional
[9] K.M.F. Fuhad, J.F. Tuba, M.R.A. Sarker, S. Momen, N. Mohammed, T. Rahman,
neural networks via concatenated rectified linear units, in: Proceedings of 33rd
Deep learning based automatic malaria parasite detection from blood smear and its
International Conference on Machine Learning (ICML2016), 2016, pp. 2217–2225,
smartphone based application, Diagnostics 10 (5) (2020) 329, https://doi.org/
arXiv Preprint arXiv: 1603.05201.
10.3390/diagnostics10050329.
[31] K. Simonyan, A. Zisserman, Very Deep Convolutional Networks for Large-Scale
[10] M. Hayat, M. Tahir, F.K. Alarfaj, R. Alturki, F. Gazzawe, NLP-BCH-Ens: NLP-based
Image Recognition, 2014 arXiv Preprint arXiv:1409.1556.
intelligent computational model for discrimination of malaria parasite, Comput.
[32] S.Y. Sourab, M.A. Kabir, A comparison of hybrid deep learning models for
Biol. Med. 149 (2022), https://doi.org/10.1016/j.compbiomed.2022.105962.
pneumonia diagnosis from chest radiograms, Sensors International 3 (2022),
[11] K. He, X. Zhang, S. Ren, J. Sun, Identity mappings in deep residual networks, in:
https://doi.org/10.1016/j.sintl.2022.100167.
European Conference on Computer Vision, 2016, pp. 630–645, arXiv Preprint arXiv:
[33] B. Sunarko, Djuniadi, M. Bottema, N. Iksan, K.A.N. Hudaya, M.S. Hanif, Red blood
1603.05027.
cell classification on thin blood smear images for malaria diagnosis, J. Phys. Conf.
[12] G. Huang, Z. Liu, L. van der Maaten, K.Q. Weinberger, Densely connected
1444 (1) (2020), https://doi.org/10.1088/1742-6596/1444/1/012036.
convolutional networks, in: Proceedings of the IEEE Conference on Computer
[34] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
Vision and Pattern Recognition, 2017, pp. 4700–4708. https://github.com/liuzh
V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in: Proceedings of
uang13/DenseNet.
the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 1–9.
[13] S. Ioffe, C. Szegedy, Batch normalization: accelerating deep network training by
[35] O.S. Zhao, N. Kolluri, A. Anand, N. Chu, R. Bhavaraju, A. Ojha, S. Tiku, D. Nguyen,
reducing internal covariate shift, in: 32nd International Conference on Machine
R. Chen, A. Morales, D. Valliappan, J.P. Patel, K. Nguyen, Convolutional neural
Learning, PMLR, 2015, pp. 448–456, in: https://proceedings.mlr.press/v37/ioffe
networks to automate the screening of malaria in low-resource countries, PeerJ 8
15.html.
(2020), https://doi.org/10.7717/peerj.9674.
[14] D.P. Kingma, J.L. Ba, Adam: A Method for Stochastic Optimization, 2014 arXiv
Preprint arXiv:1412.6980.

11