Gradio

Upload Blood Smear Image for Analysis

Upload a microscopy image of a blood cell to detect malaria parasites using AI.

Upload Blood Cell Image

Show Grad-CAM Visualization (Explainable AI)

Grad-CAM Heat Map (Where AI Looks)

Confidence Distribution

Validate Model Performance

Upload a ZIP file containing a validation dataset (with Parasitized/ and Uninfected/ folders).

Confusion Matrix

Visualize Training Metrics

View training progress, validation accuracy, and energy savings from Adaptive Sparse Training (AST).

Metrics are automatically loaded from checkpoints/metrics.csv. Upload a different file if needed.

Training Summary

Total Epochs: 30 Best Validation Accuracy: 9462.99% (Epoch 27) Final Training Loss: 0.5702 Average Energy Savings: 8811.9%

Last 5 Epochs:

epoch	timestamp	train_loss	val_acc	lr	activation_rate	energy_savings	samples_processed	total_samples	act_rate	save_rate
26	1.7628e+09	0.598898	93.8861	0.0003	0.0939399	88.8626	22046	22046	0.0939399	88.8626
27	1.7628e+09	0.581468	94.6299	0.0003	0.0939853	88.8935	22046	22046	0.0939853	88.8935
28	1.7628e+09	0.579731	94.3759	0.0003	0.0938039	88.9227	22046	22046	0.0938039	88.9227
29	1.7628e+09	0.580671	94.1945	0.0003	0.0938492	88.9499	22046	22046	0.0938492	88.9499
30	1.7628e+09	0.570211	93.9405	0.0003	0.0937585	88.9755	22046	22046	0.0937585	88.9755

Training Loss

Validation Accuracy

Activation Rate (AST)

Energy Savings (AST)

Activation Threshold (AST)

Compare Multiple Training Runs

Loss Comparison

Accuracy Comparison

Activation Comparison

Savings Comparison

Threshold Comparison

Export Model to ONNX Format

Convert the PyTorch model to ONNX format for production deployment and cross-platform compatibility.

About This System

Technology Stack

Deep Learning Framework: PyTorch
Model Architecture: EfficientNet-B0
Training Method: Adaptive Sparse Training (AST) with Sundew Algorithm
Explainable AI: Grad-CAM (Gradient-weighted Class Activation Mapping)
Dataset: NIH Malaria Cell Images (27,558 samples)

Performance Metrics

Validation Accuracy: 93.94% (final epoch), 94.63% (best epoch)
Energy Savings: 88% reduction in training cost vs. traditional methods
Inference Speed: <1 second per image
Model Size: ~16MB
Training: 30 epochs on NIH Malaria Dataset

Key Features

Real-time Diagnosis: Upload blood smear images for instant analysis
Explainable AI: Grad-CAM shows exactly where the model detects parasites
Energy Efficient: Trained using Adaptive Sparse Training with Sundew algorithm for 88% energy savings
Clinical Recommendations: Actionable advice based on predictions
Model Validation: Built-in tools for performance evaluation
ONNX Export: Deploy anywhere with standard model format

Use Cases

Research: Academic studies on malaria detection
Education: Teaching AI applications in healthcare
Triage: Rapid pre-screening in resource-limited settings
Model Comparison: Benchmark against other approaches

Important Disclaimers

⚠️ This is a research prototype and NOT a medical device.

Results must be confirmed by certified laboratory testing
Do not use for clinical diagnosis without professional validation
Always consult qualified healthcare providers
This tool is for research and educational purposes only

Developer

Oluwafemi Idiakhoa

Citation

If you use this system in your research, please cite:

@software{malaria_ast_detection,
  author = {Idiakhoa, Oluwafemi},
  title = {Malaria Detection using Adaptive Sparse Training},
  year = {2025},
  url = {https://github.com/oluwafemidiakhoa/Malaria}
}

Resources

GitHub Repository
AST Library (PyPI): pypi.org/project/adaptive-sparse-training
NIH Malaria Dataset

Built with EfficientNet-B0 + Adaptive Sparse Training | Powered by PyTorch & Gradio

🔬 AI-Powered Malaria Detection System

System Information