This is something I see firsthand. I’m an ophthalmologist working at a private medical college hospital in rural India. Every day, I encounter diabetic patients who arrive late — not because they ignored symptoms, but because the system made it nearly impossible to catch their condition early.

When my AI in Healthcare course posed the question — “AI now outperforms humans in some medical imaging tasks. Should we still require human review?” — I didn’t have to think very hard about which direction I wanted to take this.

Because for me, the more pressing question isn’t whether AI is accurate enough. It’s: accurate for whom, and available to whom?

A lot of the global conversation about AI in medical imaging is framed around urban, well-resourced hospitals. But let me paint a different picture.

India is considered the diabetic capital of the world. And a farmer from a village — who is at risk of diabetic retinopathy — faces a reality that looks nothing like a metropolitan clinic.

India doesn’t have an absolute shortage of eye specialists — it has a distribution problem. Most specialists cluster in cities, while rural populations travel hours to access tertiary care, often only when disease is already advanced.

Rural healthcare access collapses around three barriers:

By the time a patient reaches a specialist, they’re often already in advanced stages of disease. This is the gap AI has the power to address — not in a gleaming hospital, but at the front lines, during outreach camps in villages.

Here’s where I get genuinely excited. A handheld retinal scanner integrated with an AI tool capable of grading Diabetic Retinopathy — used by a trained healthcare technician at a rural outreach camp — could genuinely change lives.

CNNs (Convolutional Neural Networks) are particularly good at identifying microaneurysms in retinal vessels — tiny early signs of diabetic damage that a human eye, especially under poor lighting conditions, can miss. In a screening context, this matters enormously.

Being enthusiastic about AI doesn’t mean being uncritical. There are real, significant concerns that need to be addressed before we deploy AI tools in high-stakes screening.

• 📊

  Dataset BiasAI tools trained predominantly on Western populations may perform poorly on Indian patients. Dense cataracts — common in India — produce poor-quality retinal images that can skew results. Diverse, locally representative training data is non-negotiable.

• 🧩

  The Black Box ProblemWhen AI makes a diagnosis, doctors need to understand how it arrived there. Without explainability, trust erodes — and appropriately so. XAI (Explainable AI) frameworks are essential.

• ⚖️

  False Positives & False NegativesEven a highly accurate system makes errors. In screening contexts, a false negative (missing true disease) can have devastating consequences. Human oversight remains essential for safety.

• 🗣️

  Unqualified InterpretationAn AI diagnosis of early DR delivered to a patient without any clinical context — without a doctor to explain it — can cause unnecessary panic or dangerous complacency. AI outputs must be mediated by trained professionals.

• 📋

  Liability & AccountabilityWhen AI is the frontline diagnostician, who is responsible when something goes wrong? Clear frameworks for liability must be established — and the doctor must remain accountable.

There’s a term used in AI ethics — HITL: Human-in-the-Loop. It means the human isn’t removed from the diagnostic process; they are a critical checkpoint within it.

In the context of rural retinal screening, this might look like:

AI scans and grades the image → community health worker flags high-risk patients → images are transmitted digitally to a specialist → specialist reviews and confirms → patient receives a call with appropriate guidance and referral.

The AI handles the volume. The specialist handles the complexity and the accountability. The health worker handles the human connection on the ground. Each plays a role that the others cannot.

Machine learning also means these systems get better over time — continuously improving as they learn from larger, more diverse annotated datasets. That’s a compelling argument for investing in building robust Indian retinal imaging datasets now.