Skip to main content
New: Get an independent second opinion on your scan Try SecondRead
Artificial Intelligence 2025-03-28

Autonomous AI for Multi-Pathology Detection in Chest X-Rays

A multi-site study evaluating autonomous AI systems for detecting multiple pathologies in chest X-rays across the Indian healthcare system, demonstrating scalable diagnostic accuracy in resource-constrained environments.

5C Network Research Team · arXiv · DOI: 10.48550/arXiv.2504.00022

Key findings

  • Trained on more than 5 million chest X-rays sourced from healthcare systems across India
  • Up to 98% precision and over 95% recall for multi-pathology classification, stable across age, gender and equipment subgroups
  • Normal vs abnormal classification reached 99.8% precision, 99.6% recall and 99.9% negative predictive value
  • Deployed in 17 major Indian healthcare systems, processing over 150,000 scans at an average of 2,000 chest X-rays a day

Plain-English summary

The question
Can one AI system read chest X-rays on its own — separating normal from abnormal studies and classifying, locating and outlining a wide range of pathologies — reliably enough to work across India's very different hospitals?
The data
More than 5 million chest X-rays from healthcare systems across India. The study design includes subgroup analyses by age, gender and equipment type.
The method
The system combines Vision Transformers for classification, Faster R-CNN for detection, and several U-Net variants (Attention U-Net, U-Net++ and Dense U-Net) for segmentation.
The result
Up to 98% precision and over 95% recall for multi-pathology classification. For normal vs abnormal, 99.8% precision, 99.6% recall and 99.9% negative predictive value. Across 17 healthcare systems — diagnostic centers, large hospitals and government hospitals — it processed over 150,000 scans, and the authors report reduced reporting times.

Why it matters for hospitals and radiologists

The results come from deployment in 17 healthcare systems, not only a test set — the kind of evidence a radiology department needs before letting AI help with a high-volume study like the chest X-ray. A reliable normal/abnormal split (99.9% negative predictive value in this study) is what lets abnormal studies be prioritized for the radiologist, who still signs the report.

Related at 5C: Clinical evidence: chest X-ray AI

Citation

Bargava Subramanian, Shajeev Jaikumar, Praveen Shastry, Naveen Kumarasami, Kalyan Sivasailam, Anandakumar D, Keerthana R, Mounigasri M, Kishore Prasath Venkatesh. Autonomous AI for Multi-Pathology Detection in Chest X-Rays: A Multi-Site Study in the Indian Healthcare System. arXiv preprint arXiv:2504.00022, 2025. https://doi.org/10.48550/arXiv.2504.00022

Read Full Paper on arXiv

Related research

From the lab to the worklist

This research runs in production.

The methods described here are part of Bionic, the AI layer that pre-reads every study 5C reports — 20,000+ scans a day across 2,000+ hospitals, each one signed by an NMC-registered radiologist.