Malaria Parasite Detection in Blood Cell Images using CNN

Malaria diagnosis still depends on trained microscopists examining Giemsa-stained blood smears — slow, scarce in rural clinics, and subject to human error. Automated cell classification is one of the clearest demonstrations of CNNs doing genuinely useful pattern recognition: the visual difference between a parasitized and an uninfected cell is subtle, and a network must learn chromatin dots, pigment and morphology rather than memorizing brightness. Many student medical-AI projects train on a few hundred images and report suspiciously perfect accuracy. This project does the opposite: it trains on the full public NIH/NLM malaria dataset (27,558 single-cell images, balanced classes), uses a documented custom CNN with augmentation and early stopping, and evaluates once on a held-out test set. An interactive demo lets you run the documented pipeline on real stained blood-smear micrographs and see the prediction with its probability.

1. The 27,558 NIH cell images are loaded and split into stratified 80% train / 10% validation / 10% test sets.
2. Images are resized to 128×128, rescaled to [0, 1], and augmented during training (rotations, flips, zooms) so the network learns morphology, not orientation.
3. The custom CNN — 4 × (Conv-ReLU-BatchNorm-MaxPool) blocks → GlobalAveragePooling → Dense(128, ReLU) → Dropout(0.4) → Dense(1, sigmoid) — is compiled with Adam and binary cross-entropy.
4. Training runs with early stopping on validation loss; the best weights are checkpointed.
5. The held-out test set is evaluated once: accuracy, sensitivity, specificity and the confusion matrix are generated for the report.
6. In the web demo, selecting a sample smear field runs it through the documented preprocessing and classification pipeline and displays the verdict with its probability.

Tech stack:

• Python 3, TensorFlow/Keras
• NumPy, Matplotlib, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML5 + CSS3 + JavaScript (interactive classification demo)
• NIH/NLM malaria cell-images dataset (Rajaraman et al., 2018)

Dataset & model details

• Dataset: NIH/NLM malaria dataset (Rajaraman et al., PeerJ 2018) — 27,558 segmented single-cell images from Giemsa-stained thin blood smears, 13,779 parasitized and 13,779 uninfected, collected at Chittagong Medical College Hospital, Bangladesh.
• Task: Binary image classification; input = 128×128×3 cell image, output = P(parasitized).
• Model: Custom CNN: 4 × (Conv2D-ReLU-BatchNorm-MaxPool) blocks → GlobalAveragePooling → Dense(128, ReLU) → Dropout(0.4) → Dense(1, sigmoid); trained with Adam and binary cross-entropy, early stopping on validation loss.
• Metrics: Test accuracy 96%+ (design target), sensitivity, specificity, confusion matrix — measured once on the held-out 10% test set after the training run. No accuracy is claimed as measured until the training run is executed for the order.

• [NIH malaria dataset pipeline] Loads the real 27,558-image dataset (13,779 parasitized / 13,779 uninfected single-cell images), with documented stratified 80/10/10 splits.
• [Custom CNN architecture] Four convolution–pooling blocks with batch normalization, global average pooling, a 128-unit dense layer with dropout, and a sigmoid output — designed for this task, not copied from a tutorial.
• [Augmentation & early stopping] Rotation, flip and zoom augmentation plus early stopping on validation loss, documented as the overfitting controls.
• [Interactive classification demo] Select from six real stained blood-smear micrographs; the demo runs the documented pipeline and shows the verdict with P(parasitized) and the morphological cue behind it.
• [Full evaluation] Test accuracy, sensitivity, specificity and confusion matrix computed once on the held-out set — the metrics a medical-imaging viva actually asks about.
• [Training curves] Accuracy and loss plots for train vs validation with interpretation notes in the report.
• [Educational-scope framing] The report and demo state plainly that this is a learning prototype, not a medical device — honest scope that examiners respect.

• Complete training & evaluation Jupyter notebook
• Trained CNN model file with preprocessing code
• Interactive classification web demo with real sample micrographs
• Confusion matrix, sensitivity/specificity report and training-curve plots
• Project report PDF (background, CNN theory, dataset description, methodology, results, educational-scope statement)
• PPT presentation for final review
• Viva Q&A preparation document (convolutions, augmentation, sensitivity vs specificity, overfitting, dataset bias)