Malaria Detection from Blood Cell Images using CNN

Executive Summary
This project demonstrates how deep learning can accelerate malaria diagnosis by automatically classifying red blood cell images as Parasitized or Uninfected. Using ~27,500 labeled microscopy tiles, I benchmarked multiple CNN architectures and achieved 92% test accuracy with a VGG16 transfer learning model, maintaining high recall (93%) for parasitized cells — a clinical priority. The pipeline is designed for reproducibility, deployment feasibility, and ethical transparency, with model cards, CLI inference, and GitHub Actions CI included. This work was completed as part of the MIT Applied Data Science Program Capstone (Aug 2025).

---

📚 Table of Contents

• Dataset Overview
• Models Trained
• Results Summary
• Training Performance
• Evaluation Results
• CLI Inference
• Reproducible Environment
• Continuous Integration
• Model Card
• Impact & Risks
• Repository Structure
• Author
• License

---

Dataset Overview

• Source: NIH Malaria Dataset (27,558 thin smear RBC images)
• Splits: ~24,958 train / 2,600 test images
• Classes: Parasitized vs Uninfected
• Input: 64×64×3 tiles
• Preprocessing: RGB + HSV conversions, normalization, augmentation

Parasitized	Uninfected

Class Balance:

---

Models Trained

• Base CNN: baseline, mild overfitting mid-training
• Model 1: deeper CNN, higher peak, earlier overfit
• Model 2: BN + LeakyReLU, faster/smoother convergence
• Model 3: augmentation, improved robustness on held-out test
• VGG16 Transfer Learning: frozen backbone + dense head, fine-tuned top block

---

Results Summary

Model	Test Accuracy	Parasitized Recall	Precision	AUC (est.)	Notes
Base CNN	~0.65	~0.55	~0.60	~0.70	Baseline, mild overfit mid-training
Model 1	~0.93	~0.95	~0.92	~0.95	Deeper CNN, higher peak, earlier overfit
Model 2	~0.94	~0.96	~0.93	~0.96	BN + LeakyReLU, faster/smoother convergence
Model 3	~0.90	~0.80	~0.88	~0.88	Augmented, more robust but lower recall
VGG16	0.92	0.93	0.90	~0.94	Balanced precision/recall, best confusion matrix

---

Training Performance

Individual Models

• Model 1:
• Model 2:
• Model 3:
• VGG16:

All Models

---

Evaluation Results

Confusion Matrices

• Model 2:
• Model 3:
• VGG16:

All Models Comparison:

ROC & Recall

---

Quick CLI Inference

python src/predict.py path/to/image.png
# or
MODEL_PATH=models/malaria_vgg16.h5 python src/predict.py path/to/image.png

Reproducible Environment

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-lock.txt

CI

Minimal GitHub Actions workflow at .github/workflows/ci.yml runs ruff + smoke test.

Model Card

See MODEL_CARD.md.

---

Repository Structure

Malaria-Detection-CNN-Capstone/ │ ├── README.md # Project overview and documentation ├── LICENSE # MIT license ├── .gitignore # Ignore rules (data, models, checkpoints) ├── requirements-lock.txt # Pinned Python dependencies ├── notebooks/ │ └── Capstone_Malaria_Detection.ipynb # Main notebook (EDA + training) ├── src/ │ ├── train_model.py # Training pipeline (placeholder/extendable) │ └── predict.py # CLI inference script ├── models/ │ └── .gitkeep # Placeholder (trained models go here, e.g. malaria_vgg16.h5) ├── Images/ # Visuals for README and analysis │ ├── Data_Balance_Checks.png │ ├── Model1_Accuracy-Validation_Curves.png │ ├── Model2_Accuracy-Validation_Curves.png │ ├── Model3_Accuracy-Validation_Curves.png │ ├── VGG16model_Accuracy-Validation_Curves.png │ ├── AllComparison_Accuracy-Validation_Curves.png │ ├── Model2_ConfusionMatrix.png │ ├── Model3_ConfusionMatrix.png │ ├── VGG16model_ConfusionMatrix.png │ ├── AllComparison_ConfusionMatrix.png │ ├── AllComparison_ROC_Curve.png │ └── AllComparison_Recall.png ├── .github/ │ └── workflows/ │ └── ci.yml # GitHub Actions workflow (lint + smoke test) └── MODEL_CARD.md # Documentation of data, metrics, risks, ethics

---

How to Run

1. Install requirements:

pip install -r requirements.txt

2. Run the notebook:

open notebooks/Capstone_Malaria_Detection.html

Dataset

Download from Kaggle if not included.

Author

Gabrielle Foster – LinkedIn

⚠Research/educational purposes only. Not a clinical diagnostic tool.