MERIT AI Research

MERIT AI

Medical Embedding & Reasoning Intelligence TechnologyAn eye ultrasound is a grey image, and reading one takes years of practice. MERIT AI is being built to help a clinician read it — in three stages, each a complete piece of work in its own right, and each proven before the next is built on it. It is not a diagnostic device.

How it is built

Three Stages

Each stage answers a different question about the same scan. We say what each one does only once it does it.

Stage one

Classification

OCS · Ocular Classification System

Is there something abnormal in this eye?

The first stage looks at the scan as a whole and sorts it: abnormal, or not. It makes no attempt to say what the abnormality is, or where — only that the scan warrants a closer look. For a clinic where nobody is trained to read a B-scan, that first sort is the difference between a case that gets escalated and one that does not.

This stage is built and running.

Stage two

Segmentation

AAS · Advanced Annotation System

Where are the structures, and what shape are they?

The second stage finds the eye itself within the scan and traces its outline, separating the globe from everything around it. Where classification treats the image as one thing, segmentation understands it as a set of structures with boundaries.

This stage is live and in daily use by our validating ophthalmologist. Every result is reviewed and signed by a clinician before it counts.

Stage three

Tissue Identification

TIS · Tissue Identification System

What is each region made of?

The third stage will be the most detailed: labelling regions of the scan by tissue type — fibrous, vascular, retina, muscle, fat, gland, nerve, fluid — so the image is not just outlined but read. This is what will turn a grey picture into something a clinician who is not an ultrasound specialist can interpret.

This stage is in development.

MERIT AI is all three stages working on the same scan: classify, segment, identify. We are building them in order. A clinician reviews and signs every result.

The Problem

A Crisis of Access

Rural India has far fewer ophthalmologists than it needs, and some of the conditions that cause blindness give only a short window for treatment.

1 ophthalmologist per 70,000 rural residentsUltrasound is portable and cheapReading it well is the bottleneck

A handheld ultrasound probe can reach a district clinic long before a specialist can. What it cannot bring with it is the experience to place callipers the same way every time. MERIT AI starts there: repeatable measurement first, with a clinician deciding what the numbers mean.

Gloved hands holding an animal eye under lamp light during laboratory dissection

Animal-eye specimen prepared for scanning.

Gloved hands holding a small specimen in the laboratory

Handling under gloves, on the bench.

Our Solution

MERIT AI, On Our Own Hardware

Scans from a handheld probe are de-identified and analysed on dedicated servers we run ourselves. The output is a set of measurements for a clinician to review, not a diagnosis.

Accepts exported scans from handheld probes such as Butterfly iQ3, as PNG, JPEG or DICOM
Models run on our own GPU servers, not a third-party AI service
Three stages: classification, annotation and biometry, tissue identification
Every result reviewed and signed by a clinician
Remote specialist review planned for a later version

Dr. Hadi's Framework

The AIID Research Framework

Application → Implementation → Integration → Dissemination. The four-stage translational philosophy that drives every decision in MERIT AI's development.

Application

Start from a clinical problem: ocular ultrasound is widely available, but reading and measuring a B-scan consistently takes training that many first-contact clinics lack.

Time-critical conditions such as retinal detachment, and routine needs such as IOL planning before cataract surgery.

Implementation

Research, phantom lab training, animal testing, and data collection. Stage 1 (classification) built on phantom and synthetic data; Stage 2 (annotation and biometry) validated on phantoms and animal scans, human-eye validation under way.

175 phantom images, 600 frames/scan, 10,000-image collection target, De Cure patient pipeline.

Integration

Software development, on-premise deployment, and clinical workflow embedding. Models run on our own GPU hardware, not a public cloud, with data handling designed around HIPAA and the DPDP Act.

Containerised stack, role-based access, dual clinician sign-off on validation cases.

Dissemination

Putting the software in front of clinicians: a trial with a small group of Bengaluru doctors first, then district hospitals. Remote specialist review is planned, not yet built. NIH SBIR application in preparation.

Athreya Inc. (US) as lead, Validus Institute as co-applicant.

Research Focus

What We Are Working On

Active investigation across ophthalmic imaging, biometry, and clinical AI deployment. Peer-reviewed output is listed here once it carries a resolvable DOI.

Automated Ocular Biometry

Active

Deriving axial length and related biometric measurements from B-scan ultrasound without manual calliper placement, so that measurements are reproducible between operators rather than dependent on individual technique.

U-Net globe localisationDeterministic meridian extractionOperator-independent measurement

Optic Nerve Sheath Diameter Assessment

Active

Standardising how optic nerve sheath diameter is measured on ocular ultrasound, including where along the nerve the measurement is taken — a known source of disagreement between studies.

Standardised measurement depthInter-observer variability analysis

Portable Point-of-Care Ultrasound

Active

Evaluating handheld ultrasound for orbital assessment outside tertiary centres, where access to conventional imaging is limited. Focus on what portable hardware can and cannot reliably resolve.

Handheld probe evaluationStore-and-forward specialist review (planned)Low-resource deployment

Phantom Lab & Synthetic Data

Active

Building physical eye phantoms and synthetic B-scan datasets so that algorithms can be developed and stress-tested against known ground truth before any clinical data is involved.

Physical eye phantomsSynthetic B-scan generationGround-truth validation

On-Premises Clinical AI

Active

Running analysis models on our own hardware rather than sending patient imaging to third-party infrastructure, so that data residency and DPDPA obligations are met by architecture rather than by policy.

On-premises inferenceData residency by designDPDPA-aligned deployment

Clinician-Supervised Decision Support

Exploratory

Positioning model output as decision support reviewed by an ophthalmologist, including how confidence is surfaced so that a clinician can tell when the system is uncertain.

Confidence reportingOphthalmologist-in-the-loop review

In the laboratory

Animal-Eye and Phantom Work

How the imaging pipeline is developed and tested before it goes near a patient.

These photographs show animal-eye laboratory work and phantom preparation, used to develop and test the imaging pipeline. They are not patient imaging.

Laboratory dissection of an animal eye with instruments — Dissection under magnification, to relate what the scan shows to the tissue itself.

Animal eye on the bench during laboratory work — The specimen is measured directly, giving a known answer to check the software against.

Researcher in the phantom laboratory with materials laid out on the bench — The phantom laboratory, where eye phantoms are made and scanned.

Researcher at the laboratory bench with phantom materials — Phantom materials are built to known dimensions before any scanning.

Interested in Collaborating?

We welcome research partnerships, clinical data contributions, and institutional collaborations that move MERIT AI through external validation on human eyes.