Malaria Cell Detection

CNN that classifies red blood cells as Parasitized (malaria-infected) or Uninfected using microscope images from the NIH dataset.

Accuracy: 95.43% on held-out test data

What It Does

Takes a 64x64 microscope image of a blood cell and predicts whether the Plasmodium parasite is present. Trained on ~27,500 images (balanced classes) from the National Institutes of Health.

Model

3-block CNN built with TensorFlow/Keras:

• Block 1: 32 filters (3x3) → ReLU → MaxPool — picks up edges and color spots
• Block 2: 64 filters (3x3) → ReLU → MaxPool — combines into shapes
• Block 3: 128 filters (3x3) → ReLU → MaxPool — recognizes parasite-like structures

Then a dense layer (128 units) with 50% dropout, and a sigmoid output for binary classification.

Trained for 10 epochs with Adam optimizer and binary cross-entropy loss. 80/20 train-test split, stratified to keep class balance.

Results

• Test accuracy: 95.43%
• Generates training_history.png (accuracy/loss curves) and predictions.png (sample predictions with color-coded correctness)

Usage

Run on Kaggle with GPU enabled:

python malaria_detection.py

It auto-detects the Kaggle dataset path. For local use, put the cell_images/ folder (with Parasitized/ and Uninfected/ subfolders) in the same directory.

Dataset

NIH Malaria Cell Images

• 27,558 images total
• 13,779 Parasitized + 13,779 Uninfected
• Thin blood smear slides, stained and photographed under microscope

What I Learned

• How CNNs extract features hierarchically (edges → shapes → complex patterns)
• Why stratified splits matter for balanced evaluation
• The effect of dropout on reducing overfitting in small-ish datasets
• Image preprocessing and normalization for neural networks