You are on page 1of 5

TABLE I

Analysis of literature performed on detection of malaria parasites using artificial intelligence


approaches.

Author& Approach Focused Area Performance Limitations


year metric used

Nakasi et al. Fast R-CNN Detection of Precision and If the detection


&2020 [10] SSD, RetinaNet malaria recall confidence is
parasite from below 60%,
thick blood results in detection
smear image challenges
Fuhad et al & Automated Detection of Accuracy, Not implemented
2020 [11] convolutional malaria from Specificity, in resource-
neural network microscopic Sensitivity, constrained areas
model blood smear Precision, F1
images score
Masud et al. Custom CNN Detection of Accuracy, AUC, Not considered
& 2020 [13] model malaria Precision, data augmentation
parasite using Recall, F1Score, to avoid the
mobile MCC overfitting
application problem
Yang et al. & Custom CNN Detection of Accuracy, Does not perform
2019 [12] model malaria Sensitivity, well for very
parasite from Specificity, small objects like
thick blood Precision, parasites, with an
smears using F1Score, AUC average size of
smartphones 44x44 pixels in an
image of
4032×3024 pixels
Rahman et al. Custom CNN Detection of Accuracy, ----
& 2019 [15] model, TL- malaria from Precision,
VGG16, microscopic Recall, F1 score,
CNNEx-SVM blood smear AUC, MCC
images
Rajaraman et AlexNet, Detection of Accuracy, Only pilot studies
al. & 2018 VGG16, ResNet malaria from Sensitivity, performed in
[14] 50, Xception, thin blood Specificity, deploying models
DenseNet 121. smears Precision, into mobile
F1Score, MCC devices
Gopakumar et CNN on focus Automatic Sensitivity, ----
al. & 2020 stack detection of specificity,
[16] malaria from MCC
blood smears
TABLE II
Training Parameters used in proposed ensemble model.

Training Parameters Ensemble Methodology Cosine Annealing

Learning Rate 0.0001 0.002

Batch Size 32 32

Optimizer Adam Adam

Loss function Categorical Cross-entropy Categorical Cross-entropy

Epochs 50 130

Flipping Horizontal Horizontal

Zoom Range 15% 15%

Rotation 10 degrees 10 degrees

Lighting 20% 20%

Re-scale 1/255 1/255

TABLE III
Illustration of proposed CBRM model architecture

Strid Kernel
Layer Type Output Shape Parameters e Padding Size Dropout Filter

Input (3, 224, 224) 0 - - - 0 -

Conv2d (96, 110, 110) 7,296 2 0 5X5 0 96

BatchNorm2d (96, 110, 110) 192 - - - 0 -

ReLU (96, 110, 110) 0 - - - 0 -

MaxPool2d (96, 36, 36) 0 3 0 3X3 0 -

Conv2d (256, 38, 38) 221,440 1 2 3X3 0 256

BatchNorm2d (256, 38, 38) 512 - - - 0 -

. . . . . . .

. . . . . . .
ReLU 64 0 - - - 0 -

BatchNorm1d 64 128 - - - 0 -

Dropout 64 0 - - - 0.5 -

Softmax 2 195 - - 0 2

TABLE IV
Parameter Comparison of Proposed CBRM ensemble Cosine Annealing Model.

Outpu
t
Layer Parameter Input Layer
Model s s Layer size Size

(224,224,3
ResNet-101 101 44 X 106 ) (2,1)

(224,224,3
VGG-19 19 143 X 10 6
) (2,1)

DenseNet12 (224,224,3
1 121 8 X 10 6
) (2,1)

Proposed (224,224,3
CBR Model 6 .8 X 10 6
) (2,1)

TABLE V
Dataset Description.
Number
Class of Images
Parasitized 8658
Uninfecte
d 7904
TOTAL 16562

TABLE VI
Testing accuracy (in %) achieved by the proposed Ensemble models trained with different
batch sizes.

Batch Specificit
Size Sensitivity y Precision F1 Score Accuracy
16 96.81 97.58 97.72 97.26 97.18

32 98.31 97.69 97.81 98.06 98.01

64 97.41 97.35 97.49 97.45 97.38

TABLE VII
Classification performance (in %) assessment between various optimizers.

Optimize F1
r Sensitivity Specificity Precision Score Accuracy

RMSProp 96.91 97.3 97.45 97.18 97.1

AdaGrad 98.21 97.18 97.29 97.75 97.71

SGD 96.57 96.13 96.31 96.44 96.36

Adam 98.31 97.69 97.81 98.06 98.01

TABLE VIII
Classification Results (in %) Comparison of proposed model with other Pre-trained CNN
models.

Model Sensitivity Specificity Precision F1 Score Accuracy AUROC

VGG-19 98.71 99.13 99.17 98.94 98.91 98.8

DenseNet-
121 98.66 98.6 98.66 98.66 98.63 96.9

ResNet-101 98.97 98.48 98.55 98.76 98.73 97.1

Average
Ensemble 99.25 99.01 99.06 99.16 99.13 99.6

CBR
Cosine
Annealing 98.31 97.69 97.81 98.06 98.01 98.31

TABLE IX
Comparison of proposed model with current state-of-the-art deep learning (in %).

Reference Method Dataset used Number Accuracy


of images (%)

Yang et al. [12] Custom CNN Chittagong 1819 97.26


Medical
College
Hospital

NIH Malaria
Dataset
Masud et al. [13] Custom CNN acquired 27,558 97.30

Rajaraman et al. ResNet 50, Chittagong 27,558 95.9


[14] Medical
College
Hospital,

Rahman et al [15] TL-VGG16 NIH Malaria 27,558 97.77


Dataset

You might also like