AI for Radiology in India: Current State and What to Build
Radiology AI has moved from "research demos at conferences" to "deployed in some Indian hospitals." But the headlines run ahead of the reality, and most teams pitching radiology AI haven't actually shipped one. Here's the honest snapshot.
Key takeaways
- Mature: chest X-ray (TB, pneumonia), mammography, retinal imaging.
- Emerging: CT-based detection (stroke, lung nodule), MRI prioritization.
- Regulatory: CDSCO oversight is increasing; AI radiology tools need clinical validation.
- Practical wins: triage and prioritization, second-read, workflow automation, not autonomous diagnosis.
- Build patterns: integration with existing PACS/RIS matters more than the model.
Why this matters
The actual win in radiology AI today is not "AI radiologist." It's "AI assistant that prioritizes the urgent X-rays so the radiologist sees them first." That's a real, measurable workflow improvement that ships and runs.
What's working
Triage and prioritization
CXR algorithms that flag likely TB or pneumonia, moving those reads to the top of the radiologist's worklist. Reduces time-to-treatment.
Quality assurance
AI that flags missed findings as a "second read", catches what the human missed.
Workflow automation
Routing, image quality checks, technician feedback loops.
Specific high-volume conditions
CXR for TB (especially in screening programs), DR detection in diabetic retinopathy screening, mammography.
What's not yet ready
Autonomous diagnosis
No regulator allows AI to issue diagnostic reports without radiologist sign-off in India.
Rare-finding detection
Algorithms trained on common conditions miss the unusual ones, which are exactly the ones that matter.
Multi-modal reasoning
Combining imaging + EHR + labs into integrated diagnosis is research-stage.
Regulatory state
CDSCO classifies AI medical devices under Software as Medical Device (SaMD). Class C/D devices (higher risk) require formal approval and clinical evidence. Class A/B (lower risk) have lighter requirements.
If you're building anything that touches diagnostic decisions, plan for clinical validation studies. Skipping this is illegal and dangerous.
What to build
PACS integration
The model isn't the hard part. Integrating with hospital PACS (DICOM-based image storage) and RIS (workflow systems) is. Most successful radiology AI vendors are 70% PACS/RIS integration and 30% AI.
Worklist routing
Move identified cases to the right radiologist's worklist with appropriate priority.
Outcome tracking
Log every AI flag, every radiologist override, every clinical outcome. Use this to improve the model and prove the value.
Hospital-specific tuning
Indian patient populations and imaging protocols differ from where most foundation models trained. Plan for re-training on local data.
Common pitfalls
Treating AI accuracy as the product. It's not. Workflow fit is the product.
Skipping clinical validation. It's both regulatory and operationally necessary.
Building without a PACS-friendly integration path. Hospitals won't deploy AI that doesn't fit their existing infrastructure.
What we recommend
Pick one high-volume, low-stakes use case (CXR triage, retinopathy screening). Get PACS integration right. Run clinical validation. Ship to one hospital. Iterate. Scale.
FAQs
Can we sell to US? Yes, but FDA clearance is its own substantial project.
Open-source models? CheXNet and its derivatives are starting points; not production-ready.
Cost of clinical validation? ₹50 lakh-₹2 crore depending on scope.