Kifua AI – Chest X-ray Triage Using Deep Learning Ensembles

1. Problem & Context

Why Kifua AI?

In many Kenyan and African hospitals, there are few or no radiologists on-site. Chest X-rays are often interpreted by clinical officers or general practitioners with varying levels of radiology training.

High-burden diseases like tuberculosis, bacterial pneumonia, heart failure, and chronic lung disease depend heavily on timely and accurate X-ray interpretation.

Kifua AI is designed to:

• Provide rapid, consistent triage of chest X-rays.
• Highlight potentially serious findings.
• Help non-radiologist clinicians decide which images need urgent review.

2. Technical Architecture (High-Level)

Under the Hood

• Models:

EfficientNet, ResNet, and three additional CNN/modern backbones trained on CheXpert.

• Preprocessing:

DICOM ingestion, de-identification, resizing, normalisation.

• Ensemble engine:

Orchestrates inference across models and aggregates outputs.

• Output:

Per-finding probabilities.

Triage labels derived from those probabilities.

Grad-CAM / saliency maps for interpretability.

• Deployment Options:

On-prem GPU/CPU server or edge device in radiology / imaging unit.

API integration with existing PACS or X-ray viewing systems.

3. Impact & Evaluation

What We Aim to Measure

• Sensitivity & specificity for key findings vs radiologist reports.
• Time from X-ray acquisition to initial triage.
• Changes in:
• Time to start treatment for pneumonia/TB.
• Proportion of “missed” significant findings on retrospective review.