Deep Learning for Leukemia Detection: A MobileNetV2-Based Approach for Accurate and Efficient Diagnosis

© 2023, IRJET | Impact Factor value: 8.226 | ISO 9001:2008 Certified Journal | Page 876
Deep Learning for Leukemia Detection: A MobileNetV2-Based
Approach for Accurate and Efficient Diagnosis
Lagisetty Naga Pavithra1
1Student, Dept of Computer Science, Bangalore Institute of Technology, Bengaluru, Karnataka, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract – Acute lymphoblastic leukemia (ALL) is a type of
cancer of blood and bone marrow, which is not only fatal but
also very expensive to treat. Leukemia detection atearlystage
would can save lives and money. It is very common in children.
Most leukemia in childrenistreated. Researchstudiesreported
that leukemia brings changes in white blood cells count.
Currently for initial ALL diagnosis evaluation is done
manually. This is time consuming and prone to errors. The
proposed model is based on data collected from Kaggle
dataset. The MobileNetV2 model is a lightweight model
through which I have achieved an accuracy of 98.88% on
training data and accuracy of 98.58% on testing data, with
precision of 0.986, recall of 0.9858 and F1 score of 0.9857.
Experiments were conducted on dataset containing 3256
images from 89 patients suspected of ALL, including 25
healthy individuals. Currently the dataset contains three
stages which are Early Pre-B, Pre-B and Pro-B ALL.
Key Words: Acute lymphoblastic leukemia (ALL),
MobileNetV2, blood cancer, bone marrow, deep learning,
leukemia
1.INTRODUCTION
Leukemia is cancerof blood orbone marrowwhichproduces
blood cells. It usually involves white blood cells.Whiteblood
cells are infection fighters, where they divide in orderly way
to fight as your body needs them. But when it comes to
people with leukemia there is abnormal amount of white
blood cells. Treatment is very complicated and varies based
on the type of leukemia tat the person is facing and involves
various other factors. There are many types of leukemia but
few of them are very common in children which is ALL. It
occurs in children of age 2 to 4. Acutemyelogenousleukemia
(AML) is second most common in children. ALL can affect
different types of lymphocytescalledB-cellsorT-cells.Blood
stem cells originate in the bone marrow, mainly in flatbones
in adults (hip, sternum, skull, ribs, vertebrae, scapulae, to
name a few.), and can follow two developmental lines. Cells
of the myeloid lineage give rise to white blood cells,
especially neutrophil monocytes, platelets, and red blood
cells; Cells of the lymphoid lining produce white blood cells,
also called lymphocytes [1].
Leukemia occurs when there is damage in DNA of
developing blood cells, mainlywhitebloodcellswhichcauses
the blood cells to divide and grow uncontrollable.
Researchers say that leukemia might be genetic and run in
the family.
Few of the symptoms of leukemia include fatigue,
weakness, pale skin, fever, and chills. It also causes
headaches, nausea, vomiting, confusion, seizures. Leukemia
can cause petechiae, a rash like collection of pinpoint red
spots on the skin.
Currently leukemia is not curable. In few cases it is
treatable with chemotherapy, radiation therapy, stem cell
transplantation, CART-cell therapy,targettherapyandother
methods. A risk factor is anything that may increase your
chance of having a disease. Some of the risk factors include
smoking, exposure to certain chemicals, radiation exposure
and blood disorders.
There are several methods that can be used to detect
leukemia like CNN, EfficientNet,ResNet,DenseNetandmany
more. The proposed model is based on MobileNetV2.
MobileNetV2 excels due to its efficiency, lightweight design,
and fast inference, making it suitable for resource-
constrained environments and real-time applications. It
offers strong generalization through pre-trained weights,
robust feature extraction, scalability, and competitive
performance, backed by open-source accessibility and
community support.
The rest of the paper is organized as follows: Section II
Literature review, Section III presents the proposed
architecture, and IV and V explain the experimental settings,
results, and conclusion. Finally, Section VI discusses the
future work.
2. LITERATURE REVIEW
N. Jiwani et al. [1] introduced a pioneering approach using
Pattern Recognition and Computational Deep Learning to
enhance acute lymphoblastic leukemia (ALL) diagnosis and
management. The ALLDM model achieved impressive
accuracy rates, such as 87.92% in chemotherapy
management and 94.31% in stem cell transplantation
management. This technology holds promise for improving
ALL treatment outcomes, especially in children.
A. Batool et al. [2] proposed a comprehensive solution to
address the diagnostic complexities of acute lymphoblastic
leukemia (ALL). Their work introduces a state-of-the-art DL
model based on EfficientNet-B3, achieving remarkable
accuracy in leukemia cell classification. This model
outperforms existing DL classifiers, offering a robust and
reliable tool to enhance clinical leukemia detection and
improve patient outcomes.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 10 Issue: 10 | Oct 2023 www.irjet.net p-ISSN: 2395-0072

M. A. Hossain et al. [3] proposed a cost-effective solution to
detect early-stage leukemia based on symptoms. Their
explainable supervised machine learning model, using
decision trees and the Apriori algorithm,outperformedother
algorithms with a 97.45% accuracy rate. Sharing the dataset
and code enhances resources for future leukemia research.
N. Akram et al. [4] introduced a pioneering solution for
leukemia diagnosis, focusing on WBC segmentation. Their
multi-scale information fusion network (MIF-Net), with its
internal and external spatial informationfusionmechanisms,
excelled in the accurate segmentation of challenging WBC
images. Across four datasets, MIF-Net achieved state-of-the-
art segmentation performance, boasting remarkable
accuracy, and it maintains computational efficiency with just
2.67 million trainable parameters.
Atteia, G et al. [5] introduced a Bayesian-optimized
convolutional neural network (CNN) foracutelymphoblastic
leukemia(ALL)detectioninbloodsmearimages.Themodel's
hyperparameters were tailored usingBayesianoptimization,
resulting in enhanced classification performance. This
innovative approach yielded superior accuracy,
outperforming other optimized deep learning models,
promising improved ALL detection.
Chen et al. [6] introduced aResnet101-9 ensemble model for
acute lymphoblasticleukemia(ALL)detectioninmicroscopic
images, combining nine trained Resnet-101 models with
majority voting. Algorithm hyperparameterswereoptimized
through the Taguchi method. The model achieved an
accuracy of 85.11% and an F1-score of 88.94, surpassing
individual models and excelling in precision, recall, and
specificity.
Houssein EH et al. [7] introduced an end-to-end computer-
aided diagnosis (CAD) system for leukocyte classification
using deep learning. They combined DenseNet-161 with
cyclical learning rateand the one-cycletechniquetooptimize
hyperparameters. The model achieved remarkableaccuracy,
with 100% on the training set and 99.8% on the testing set,
promising significant improvements in white blood cell
classification.
Kruse A et al. [8] proposed an advanced model for Minimal
Residual Disease (MRD) detection, crucial for predicting
leukemiarelapse.Theyleveragednext-generationsequencing
(NGS) to enhance MRD diagnostics' sensitivity. The model
employed phenotypic markers and differentialgenepatterns
analysed through various techniques like flow cytometry
(FCM), PCR, RQ-PCR, RT-PCR, or NGS.
Bibi N et al. [9] presented an IoMT-based framework for
quick and safe leukemia identification, aiming to address the
shortcomings of existing methods. Leveraging cloud
computing, this system facilitates real-time coordination for
diagnosisandtreatment.UsingDenseNet-121andResNet-34,
the study outperformed other algorithms in identifying
leukemia subtypes.
Loey et al. [10] proposed two automated leukemia
classification models using transfer learning for early
detection. The first model preprocesses images and employs
a pre-traineddeepconvolutionalneuralnetwork,AlexNet,for
feature extraction and classification. In the second model,
AlexNet is fine-tuned for improved performance.
Experiments on 2820 images demonstrated that the second
model achieved a remarkable 100% classification accuracy,
surpassing the first model.
3. PROPOSED SYSTEM
Model is designed to automate leukemia classification using
microscopic images. This system leverages deep learning
and transfer learning techniques to enhance accuracy and
efficiency.
3.1 DATASET COLLECTION
The implementation of the proposed model begins by
loading the dataset. The dataset used for this model is
available publicly on Kaggle. The dataset contains 3256
images classified as healthy, Early Pre-B, Pre-B and Pro-B.
There are 504 healthy, 985 Early Pre-B, 963 Pre-B and 804
Pro-B. It is split into training validation and testing in the
ratio 80:10:10.
3.2 DATA PREPROCESSING
The preprocessing techniques used in the code involve
resizing and rescaling images to a target size of 256x256
pixels and normalizing their pixel values to a range between
0 and 1. Additionally, data augmentation methods such as
random flips and rotations are applied to increase the
diversity of the training dataset. Cachingandprefetching are
utilized to improve data loading efficiency during training,
ensuring optimal performance of the deep learning model.
3.3 TRANSFER LEARNING
Model utilizes MobileNetV2, a pre-trained deep
convolutional neural network, as the feature extractor. This
model, trained on a large and diverse dataset, offers a
valuable starting point for leukemia classification. By
extracting high-level features from images, MobileNetV2
provides valuable insights into image content.
3.4 MODEL ARCHITECTURE
The system builds on MobileNetV2 withadditional layersfor
classification. These layers include a Global Average Pooling
2D layer to reduce dimensionality, a Dropout layer to
prevent overfitting, and fully connected Dense layers. The
final Dense layer has a softmax activation function to output
class probabilities.

3.5 TRAINING AND EVALUATION
Model is trained using the preprocessed training dataset
with a batch size of 32 for 30 epochs. The training process is
monitored for accuracyandloss.Themodel'sperformanceis
evaluated using a separate test dataset.
3.6 METRICS AND VISUALIZATION:
The system provides insights into the model's performance
by calculating metrics such as accuracy and loss. These
metrics offer a quantitative measure of the model's
classification capabilities. Additionally,thesystemgenerates
visualizations, including accuracy and loss curves over the
training epochs.
4. RESULTS AND ANALYSIS
Chart -1: Training and Validation accuracyandloss
It was observed that the proposed neural network model
achieved an accuracy of98.87%ontrainingdata and98.58%
on testing data. The loss reduced from 0.1874 to0.0443.The
highest value of accuracy was achieved at 18th epoch.
Table -1: Output for proposed model w.r.t intermediate
epochs
Epoch Loss Accuracy%
1 0.1874 93.18
15 0.0645 97.73
30 0.0443 98.58
Fig -1: Actual vs Predicted
Figure 1 shows predictions of few of the randomly selected
images from dataset.
5. CONCLUSION
In this study, a deep learning model based on MobileNetV2
demonstrated impressive accuracy (98.58%) in classifying
acute lymphoblastic leukemia (ALL) from microscopic
images. It boasted high precision and recall (both above
98%) and a strong F1 score (0.986). The model was trained
on a diverse dataset of 3256 images, encompassingdifferent
ALL stages and healthy samples. The utilization of transfer
learning with MobileNetV2 enhanced its classification
capabilities. This research offers significant potential for
early ALL detection, providing a valuable tool for medical
professionals and the possibility of improving patient
outcomes, while further refinements could advance its
clinical utility.
6. FUTURE WORK
In the future, the dataset canbeextendedbyaddingnew
samples and utilizing new augmentation techniques. A
variety of deep learning models can be applied to improve
accuracy. Various feature extraction techniques could be
used.
REFERENCES
[1] N. Jiwani, K. Gupta, G. Pau and M. Alibakhshikenari,
"Pattern Recognition of Acute Lymphoblastic Leukemia
(ALL) Using Computational Deep Learning," in IEEE

Access, vol. 11, pp. 29541-29553, 2023, doi:
10.1109/ACCESS.2023.3260065.
[2] A. Batool and Y. -C. Byun, "Lightweight EfficientNetB3
Model Based on Depthwise Separable Convolutions for
Enhancing Classification of Leukemia White Blood Cell
Images," in IEEE Access, vol. 11, pp.37203-37215,2023,
doi: 10.1109/ACCESS.2023.3266511.
[3] M. A. Hossain, A. K. M. M. Islam, S. Islam, S. Shatabda and
A. Ahmed, "Symptom Based Explainable Artificial
Intelligence Model for Leukemia Detection," in IEEE
Access, vol. 10, pp. 57283-57298, 2022, doi:
10.1109/ACCESS.2022.3176274.
[4] N. Akram et al., "Exploiting the Multiscale Information
Fusion Capabilities for Aiding the Leukemia Diagnosis
Through White Blood Cells Segmentation," in IEEE
Access, vol. 10, pp. 48747-48760, 2022, doi:
10.1109/ACCESS.2022.3171916.
[5] Atteia, G.; Alhussan, A.A.; Samee, N.A. BO-ALLCNN:
Bayesian-Based Optimized CNN for Acute
Lymphoblastic Leukemia Detection in Microscopic
Blood Smear Images. Sensors 2022, 22, 5520.
https://doi.org/10.3390/s22155520.
[6] Chen, YM., Chou, FI., Ho, WH. et al. Classifying
microscopic images as acute lymphoblastic leukemia by
Resnet ensemble model and Taguchi method. BMC
Bioinformatics 22 (Suppl 5), 615 (2021).
https://doi.org/10.1186/s12859-022-04558-5
[7] Houssein EH, Mohamed O, Abdel Samee N, Mahmoud
NF, Talaat R, Al-Hejri AM, Al-Tam RM. Using deep
DenseNet with cyclical learning rate to classify
leukocytes for leukemia identification. Front Oncol.
2023 Sep 12;13:1230434. doi:
10.3389/fonc.2023.1230434.PMID:37771437;PMCID:
PMC10523295.
[8] Kruse A, Abdel-Azim N, Kim HN, Ruan Y, Phan V, Ogana
H, Wang W, Lee R, Gang EJ, Khazal S, Kim YM. Minimal
Residual Disease Detection in Acute Lymphoblastic
Leukemia. Int J Mol Sci. 2020 Feb 5;21(3):1054. doi:
10.3390/ijms21031054. PMID: 32033444; PMCID:
PMC7037356.
[9] Bibi N, Sikandar M, Ud Din I, Almogren A, Ali S. IoMT-
Based Automated Detection and Classification of
Leukemia Using Deep Learning. J Healthc Eng. 2020 Dec
3;2020:6648574. doi: 10.1155/2020/6648574. PMID:
33343851; PMCID: PMC7732373.
[10] Loey, M.; Naman, M.; Zayed, H. Deep Transfer Learning
in Diagnosing Leukemia in Blood
Cells. Computers 2020, 9, 29.
https://doi.org/10.3390/computers9020029
BIOGRAPHIES
Lagisetty Naga Pavithra is a
student at Bangalore Institute of
Technology, Bengaluru,Karnataka,
India pursuing Computer Science
and Engineering.

Deep Learning for Leukemia Detection: A MobileNetV2-Based Approach for Accurate and Efficient Diagnosis

More Related Content

Similar to Deep Learning for Leukemia Detection: A MobileNetV2-Based Approach for Accurate and Efficient Diagnosis (20)

More from IRJET Journal (20)

Recently uploaded (20)

Deep Learning for Leukemia Detection: A MobileNetV2-Based Approach for Accurate and Efficient Diagnosis