2nd INTERNATIONAL SYMPOSIUM ON COLLABORATIVE INFORMATICS (2nd

ISCI 2025)
17–19 December 2025, University BDT College of Engineering, Davanagere,
India

Deep Learning-Based Malaria Prediction

Classifier System
Harish B G1, Srinivasulu M2, G C Pooja3, Sanjana TM4, Venkatesh N5
1. Harish B G, Research Scholar, Professor, Department of Master Of Computer
Applications; University BDT College of Engineering, Davanagere 577004, India.
Email : harishbg@[Link]
2. Srinivasulu.M Research Scholar, Professor, Department of Master Of Computer
Applications; University BDT College of Engineering, Davanagere 577004, India.
Email : srinivasulu@[Link]
3. G C Pooja PG Student, Department of Master of Computer Applications,
University BDT College of Engineering, Davanagere, Karnataka, India - 577004.
Email: poojagc1763@[Link]
4. Sanjana TM PG Student, Department of Master of Computer Applications,
University BDT College of Engineering, Davanagere, Karnataka, India - 577004.
Email: sanjanatm11@[Link]
5. Venkatesh N PG Student, Department of Master of Computer Applications,
University BDT College of Engineering, Davanagere, Karnataka, India - 577004.
Email: venkisubasv@[Link]

ABSTRACT
Malaria continues to be one of the most serious and widespread infectious diseases,
causing millions of infections and hundreds of thousands of deaths every year,
especially in resource-limited countries. Manual microscopic diagnosis using Giemsa-
stained blood smears, although accurate, is time-consuming and depends heavily on
skilled personnel. This paper proposes a deep learning-based classifier system for the
automatic detection of malaria from digital blood smear images. Six Convolutional
Neural Network (CNN) architectures—two custom-designed models (Sequential CNN
and Custom CNN) and four pre-trained models (VGG-16, DenseNet121, MobileNet V2,
and EfficientNet V2S)—were trained and evaluated using the NIH malaria dataset.
Among them, the Custom CNN achieved the highest accuracy of 96%, while
DenseNet121 and MobileNet V2 achieved 94% each. The results demonstrate that
deep learning significantly improves the accuracy and speed of malaria diagnosis. A
web-based application developed using Streamlit allows users to upload blood smear
images and receive real-time predictions. This study confirms that deep learning
provides an efficient, accessible, and cost-effective diagnostic tool to support medical
professionals in early malaria detection.
Keywords: Malaria, Deep Learning, CNN, Image Classification, Medical Diagnosis

INTRODUCTION
Malaria remains one of the most significant and persistent global health challenges of
our time. Caused by protozoan parasites of the Plasmodium genus and transmitted
through the bites of infected female Anopheles mosquitoes, the disease continues to
exact a heavy toll on human life and well-being. According to the World Health
Organization (WHO) World Malaria Report 2023, there were an estimated 249 million
malaria cases and 608,000 related deaths globally in 2022. The vast majority of this
burden is concentrated in the WHO African Region, perpetuating cycles of poverty
and placing immense strain on already fragile healthcare systems. Prompt and
accurate diagnosis remains a cornerstone of malaria control and elimination
strategies, enabling timely treatment, prevention of severe disease progression, and
reduction of transmission.
For over a century, the gold standard for malaria diagnosis has been the microscopic
examination of Giemsa-stained thick and thin blood smears. This conventional
approach allows for the direct visualization and quantification of malaria parasites,
offering crucial clinical insights [2]. However, it is labor-intensive, time-consuming,
and highly dependent on the expertise of trained microscopists, which introduces
inconsistencies—especially at low parasite densities. In resource-limited regions
where malaria is endemic, the availability of well-trained technicians and reliable
laboratory infrastructure remains severely restricted, creating a critical bottleneck in
timely patient diagnosis.
The recent and rapid advancements in Artificial Intelligence (AI), particularly within
the subfield of Deep Learning (DL), have introduced transformative possibilities for
medical image analysis. Deep learning models, specifically Convolutional Neural
Networks (CNNs), have achieved exceptional success in computer vision tasks, often
matching or surpassing human-level performance in image classification [4]. CNNs
can automatically learn spatial hierarchies of features from digital images and, when
trained on malaria blood smears, can differentiate parasitized from uninfected cells
with high precision.
This paper presents the development, comparative evaluation, and implementation
of a deep learning–based classifier system for automated malaria detection [6]. The
study investigates six CNN architectures, including two custom-built models
(Sequential CNN and Custom CNN) and four pre-trained architectures—VGG-16,
DenseNet121, MobileNet V2, and EfficientNet V2S—to evaluate their effectiveness for
this diagnostic task [7]. Furthermore, a web-based interface is developed to provide
real-time predictions from blood smear images. The primary objective is to validate
the feasibility of deep learning as a fast, reliable, and accessible diagnostic tool to
complement traditional microscopy workflows.

LITERATURE REVIEW
Over the past decade, numerous studies have focused on automating malaria
diagnosis through digital image processing and machine learning techniques. Earlier
research primarily relied on traditional image processing methods, which involved
manual feature extraction such as color, texture, and shape descriptors from
microscopic blood smear images. These features were then classified using
conventional algorithms like Support Vector Machines (SVM), Decision Trees, or k-
Nearest Neighbors (k-NN). Although such methods demonstrated moderate accuracy,
their performance was highly dependent on feature selection quality and sensitive to
variations in staining, lighting, and image acquisition conditions.
The introduction of Deep Learning (DL) and Convolutional Neural Networks (CNNs)
revolutionized medical image analysis by enabling automatic feature learning from
raw images. Rajaraman et al. (2018) demonstrated that transfer learning using pre-
trained CNN models such as VGG-16 and ResNet significantly improved malaria
classification accuracy on large-scale datasets. Dong et al. (2017) also validated the
effectiveness
of CNNs for identifying malaria parasites in thin blood smears, achieving superior
accuracy compared to traditional machine learning models. Similarly, Howard et al.
(2017) introduced MobileNet, a lightweight CNN architecture optimized for mobile
devices, making AI-assisted diagnosis feasible even in low-resource settings.
Recent studies have also explored DenseNet121 and EfficientNet architectures for
medical imaging tasks due to their efficient feature reuse and reduced computational
cost. These networks have shown promising results in achieving high classification
accuracy while maintaining minimal inference time. Despite these advancements, a
consistent limitation among existing studies is the lack of comparative evaluation
across multiple architectures under identical experimental conditions. Furthermore,
few studies have focused on the deployment of trained models into real-time
applications, which is essential for practical use in hospitals and field environments.
This research addresses these gaps by conducting a comprehensive comparison of
six CNN architectures—including custom-built and transfer learning models—and
implementing a web-based application that demonstrates the feasibility of
automated, real-time malaria diagnosis.

METHODOLOGY
The methodology followed in this study outlines the systematic development and
evaluation of a deep learning-based classifier system for automated malaria
detection. The approach involves dataset selection, image preprocessing and
augmentation, model design, training, evaluation, and deployment into a functional
web application.
 Dataset Description
The experiments were conducted using the publicly available NIH Malaria Dataset,
developed by the Lister Hill National Center for Biomedical Communications
(LHNCBC), part of the U.S. National Library of Medicine. The dataset comprises
27,558 segmented images of red blood cells (RBCs), equally divided between two
classes:
 Parasitized cells: 13,779 images infected with Plasmodium parasites.
 Uninfected cells: 13,779 healthy RBC images.
All images are color PNG files with an average resolution of 100×100 pixels. The
dataset’s balanced nature prevents model bias and ensures fair evaluation.
 Data Preprocessing and Augmentation
Before training, images were standardized through a sequence of preprocessing
steps to enhance model robustness and generalization:
 Resizing: All images were resized to 128×128 pixels.
 Normalization: Pixel values (0–255) were scaled to a [0, 1] range.
 Data Augmentation: To prevent overfitting and expand data diversity, random
transformations were applied—such as rotation (±15°), horizontal and vertical
flips, width/height shifts (10%), and zoom (10%).
The dataset was split into three mutually exclusive subsets:
 Training set: 80% (22,046 images)
 Validation set: 10% (2,756 images)
 Testing set: 10% (2,756 images)
  CNN Architectures Used
Six different CNN architectures were implemented and evaluated:
1. Sequential CNN: A simple, custom-built CNN used as a baseline.
2. Custom CNN: A modified CNN with additional convolutional layers and dropout
for regularization.
3. VGG-16: A deep network with stacked 3×3 convolutional layers, pre-trained
on ImageNet.
4. DenseNet121: A densely connected architecture allowing feature reuse and
efficient gradient propagation.
5. MobileNet V2: A lightweight, efficient model optimized for mobile and low-
resource environments.
6. EfficientNet V2S: A compound-scaled model designed to balance accuracy
and computational efficiency.
(Table 1: Summary of CNN architectures used for malaria classification)
 Model Training Setup
All models were trained using the TensorFlow and Keras frameworks. Training was
accelerated using an NVIDIA Tesla T4 GPU. The following parameters were used
consistently:
 Optimizer: Adam (learning rate = 0.001)
 Loss function: Binary cross-entropy
 Batch size: 32
 Epochs: Up to 50 (with early stopping after 5 epochs without validation
improvement)
Transfer learning was applied to VGG-16, DenseNet121, MobileNet V2, and
EfficientNet V2S. Initially, the pre-trained convolutional layers were frozen, and only
the newly added dense layers were trained. Later, partial fine-tuning was performed
with a reduced learning rate to adapt the model to malaria cell image characteristics.
 Evaluation Metrics
Model performance was evaluated on the held-out test set using standard metrics:
 Accuracy: (TP + TN) / (TP + TN + FP + FN)
 Precision: TP / (TP + FP)
 Recall (Sensitivity): TP / (TP + FN)
 F1-score: 2 × (Precision × Recall) / (Precision + Recall)
Additionally, confusion matrices were plotted to visualize the true positive (TP), true
negative (TN), false positive (FP), and false negative (FN) distributions for each
model.
 System Implementation
To demonstrate real-world feasibility, the best-performing CNN model was integrated
into a web-based application using the Streamlit framework. The system architecture
is as follows:
 1. The user uploads a cell image.
2. The backend preprocesses the image (resize and normalize).
3. The model predicts the cell’s class (parasitized or uninfected).
4. The predicted label and confidence score are displayed on the interface.

(Figure 1: Architecture of the web-based malaria prediction system)

The application also includes a lightweight authentication module and supports real-
time predictions, making it an effective proof-of-concept for clinical deployment.

ANALYSIS AND RESULTS

The performance of each CNN architecture was carefully evaluated to determine the
most efficient and reliable model for malaria cell classification. The experiments were
designed to measure not only accuracy but also the model’s ability to generalize
across unseen test data.
 Performance Evaluation
Each model was trained for a maximum of 50 epochs with early stopping enabled.
Training and validation accuracy curves confirmed stable convergence, indicating
that the models effectively learned discriminative features. Among all, the Custom
CNN demonstrated superior performance, achieving a test accuracy of 96%, while
DenseNet121 and MobileNet V2 achieved 94% each. The VGG-16 model attained
91% accuracy, while EfficientNet V2S lagged at 62%, likely due to overfitting and
computational sensitivity to smaller datasets.
(Table : Model Performance Comparison on the Test Dataset)
Model Accuracy (%) Precision Recall F1-Score

Sequential CNN 90 0.89 0.91 0.90

Custom CNN 96 0.95 0.97 0.96

VGG-16 91 0.90 0.91 0.90

DenseNet121 94 0.93 0.94 0.94

MobileNet V2 94 0.93 0.93 0.93

EfficientNet V2S 62 0.61 0.63 0.62

(Figure 2: Accuracy comparison of CNN models used for malaria detection)

 Confusion Matrix Analysis

The confusion matrices for the top-performing models (Custom CNN, DenseNet121,
and MobileNet V2) show excellent classification consistency. The Custom CNN model
produced minimal false negatives, which is crucial in medical diagnosis where
missing a positive case can have severe consequences.

(Figure 3: Confusion Matrix for the Custom CNN Model)

The results validate the reliability of deep learning in detecting malaria-infected cells
from microscopic images. The high precision and recall scores further confirm the
model’s effectiveness in distinguishing parasitized from uninfected cells.
 Comparative Discussion
Compared to traditional machine learning and manual feature-based methods, the
CNN-based approach significantly improves diagnostic performance. Dense
connections in DenseNet121 enabled better gradient flow and feature reuse, leading
to robust generalization, while MobileNet V2, due to its depthwise separable
convolutions, provided a lightweight yet powerful solution suitable for real-time
applications.
The integration of a Streamlit-based web interface allowed the model to be tested in
real-world conditions. Users could upload blood smear images, and the application
would instantly return the classification result and confidence level. This validated
the usability and scalability of the system for healthcare professionals in remote or
resource-limited settings.
  Summary of Results
The final analysis confirms that the proposed Custom CNN model is the most suitable
for automated malaria detection. Its balance between accuracy, computational cost,
and generalization capability makes it ideal for real-time deployment scenarios.

Efficient Net V2s-Confusion Matrix

FUTURE ENHANCEMENTS
While the current study successfully demonstrates the high efficacy of deep learning
models for malaria detection and presents a functional proof-of-concept application,
the research opens several promising avenues for future work. The journey from a
high-performing computational model to a clinically validated, globally deployable
diagnostic tool requires addressing challenges in data diversity, model
intelligence, and practical implementation. The following areas represent the
most critical and impactful directions for future development.

 Enhancing Model Generalizability with Diverse Datasets

The models used in this study were trained exclusively on the NIH dataset, which,
while extensive and well-curated, contains images from a single source with uniform
imaging conditions. A primary priority for future work is to improve the model’s
robustness and generalizability by training and validating it on more diverse datasets.

• Multi-Center Data: Sourcing blood smear images from different laboratories and
geographical regions to expose the model to variations in slide preparation, staining
protocols, microscope types, and lighting conditions.
• Whole Slide Imaging (WSI): Transitioning from pre-segmented cell images to full-
slide images, where the system must both detect and classify individual red blood
cells (RBCs). This represents a realistic clinical environment and transforms the
problem into an object detection challenge rather than simple classification.

 Expanding Diagnostic Capabilities: Multi-Class Classification

The current framework performs a binary classification (parasitized vs. uninfected). A
major enhancement would be to extend the model to multi-class classification,
enabling more detailed diagnostic information.
• Parasite Species Identification: Differentiating between multiple Plasmodium
species (P. falciparum, P. vivax, P. malariae, P. ovale), which is clinically essential for
guiding appropriate treatment.
• Life Cycle Staging: Identifying the parasite’s developmental stages (ring,
trophozoite, schizont) within RBCs to assess infection severity and monitor treatment
response.

 Improving Interpretability with Explainable AI (XAI)

One of the major barriers to clinical adoption of deep learning models is their black-
box nature. To gain the trust of clinicians and pathologists, it is vital to improve
transparency in model decision-making.
Future work should integrate Explainable AI (XAI) techniques such as:
• Class Activation Mapping (CAM) and Grad-CAM: Generating heatmaps that overlay
the input image to highlight regions influencing the model’s prediction.
These visual explanations allow healthcare professionals to verify that the model is
focusing on biologically relevant features, thereby increasing confidence and
improving debugging of misclassifications.

 Real-World Deployment: Edge Computing and Mobile Integration

The current web-based prototype relies on client–server communication and an
active internet connection, which can limit usability in low-resource settings. The
next logical step in deployment is to transition toward edge computing and mobile
integration.
• Model Optimization: Techniques like quantization (reducing precision from 32-bit to
8-bit) and pruning (removing redundant parameters) can drastically minimize model
size while maintaining accuracy.
• Offline Mobile Application: Developing a lightweight Android/iOS application capable
of running the optimized model locally on a smartphone, enabling offline malaria
detection at the point of care. This would empower healthcare workers to capture
images via microscope-mounted cameras and receive immediate analysis without
requiring connectivity.

 From Classification to Quantification: Automating Parasitemia Calculation

In clinical practice, malaria diagnosis also requires quantifying parasite density
(parasitemia), which is the percentage of infected red blood cells in a given sample.
Automating this process would provide valuable quantitative insight into disease
severity.
Using object detection frameworks such as YOLO or Faster R-CNN, the system could
be extended to:
1. Detect all RBCs within a blood smear image.
2. Classify each detected cell as parasitized or uninfected.
3. Automatically calculate parasitemia percentage.
This enhancement would eliminate one of the most time-consuming steps in manual
microscopy and provide standardized, rapid, and quantitative results that are
clinically relevant.

CONCLUSIONS
The proposed deep learning-based malaria prediction classifier system effectively
demonstrates the power of artificial intelligence in medical image analysis. Through a
comparative evaluation of six CNN architectures—including both custom-built and
pre- trained models such as VGG-16, DenseNet121, MobileNet V2, and EfficientNet
V2S— the study achieved a maximum accuracy of 96% using the Custom CNN model.
The results confirm that deep learning significantly improves diagnostic precision and
efficiency compared to conventional microscopy-based methods.
Furthermore, the deployment of this model within a web-based interface highlights
its potential for real-time, accessible malaria screening, especially in low-resource
settings. This research lays the groundwork for scalable AI-driven diagnostic systems
capable of assisting healthcare professionals in rapid and accurate malaria detection.

ACKNOWLEDGEMENTS
The authors express their sincere gratitude to the Department of MCA, University
BDT College of Engineering, Davanagere, for providing the facilities, technical
support, and encouragement throughout this research. The authors also wish to
thank Dr. Harish B G, Head of the Department and Associate Professor, for his
valuable guidance and mentorship during the project development. The constructive
feedback and cooperation from faculty members and peers greatly contributed to
the success of this work.
REFERENCES
[1] World Health Organization (WHO). World Malaria Report 2023. Geneva: World
Health Organization, 2023.
[2] Ross, N. E., Pritchard, C. J., Rubin, D. M., & Duse, A. G. “Automated image
analysis for malaria parasite detection in thin blood smears.” Medical and
Biological Engineering and Computing, vol. 44, no. 5, pp. 427–436, 2006.
[3] Diaz, G., Gonzalez, F. A., & Romero, E. “Automatic segmentation and
classification of human malaria parasites from thick blood smear images.”
Proceedings of the 2009 Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC), pp. 6184–6187, 2009.
[4] Dong, Y., Jiang, Z., Shen, H., & Pan, W. D. “A deep learning-based method for
malaria detection in thin blood smears.” IEEE Journal of Biomedical and Health
Informatics, vol. 21, no. 6, pp. 1633–1643, 2017.
[5] Liang, Z., Powell, A., Ersoy, I., Poostchi, M., Silamut, K., Palaniappan, K., Guo, P.,
Hossain, M. A., Sameer, A., Maude, R. J., & Jaeger, S. “CNN-based image analysis for
malaria diagnosis.” Proceedings of the IEEE International Conference on
Bioinformatics and Biomedicine (BIBM), pp. 493–496, 2017.
[6] Rajaraman, S., Antani, S., Poostchi, M., Silamut, K., Hossain, M. A., Maude, R. J.,
Jaeger, S., & Thoma, G. R. “Pre-trained convolutional neural networks as feature
extractors toward improved malaria parasite detection in thin blood smear images.”
PeerJ, vol. 6, e4568, 2018.
[7] Ahsan, M. M., Mahmud, M. P., Saha, P. K., Gupta, K. D., & Siddique, Z. “Effect of
data scaling methods on machine learning algorithms and model performance.”
Technologies, vol. 8, no. 3, 52, 2020.
[8] Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T.,
Andreetto, M., & Adam, H. “MobileNets: Efficient convolutional neural networks for
mobile vision applications.” arXiv preprint arXiv:1704.04861, 2017.
[9] Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. “Densely connected
convolutional networks.” Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), pp. 4700–4708, 2017.
[10] Tan, M., & Le, Q. V. “EfficientNetV2: Smaller models and faster training.”
Proceedings of the 38th International Conference on Machine Learning (ICML), pp.
10096–10106, 2021.
[11] Rajaraman, S., & Antani, S. “Modality-specific deep learning model ensembles
toward malaria parasite detection in thin blood smear images.” PeerJ, vol. 8,
e10294, 2020.
[12] Poostchi, M., Silamut, K., Maude, R. J., Jaeger, S., & Thoma, G. R. “Image
analysis and machine learning for detecting malaria.” Translational Research, vol.
194, pp. 36–55, 2018.
[13] Mahapatra, S., & Sahu, P. P. “A comparative study on handcrafted and deep
learning-based feature extraction for malaria parasite detection.” Biomedical
Signal Processing and Control, vol. 68, 102738, 2021.
[14] Goodfellow, I., Bengio, Y., & Courville, A. Deep Learning. MIT Press, 2016.
[15] Krizhevsky, A., Sutskever, I., & Hinton, G. E. “ImageNet classification with deep
convolutional neural networks.” Advances in Neural Information Processing Systems
(NIPS), vol. 25, pp. 1097–1105, 2012.
[16] Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., & Chen, L. C. “MobileNetV2:
Inverted residuals and linear bottlenecks.” Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), pp. 4510–4520, 2018.
[17] WHO. “Global Technical Strategy for Malaria 2021–2030.” Geneva: World Health
Organization, 2021.