AI-Assisted Eye Pathology Screening and Recommendation Systems

Summary

We build assisted eye screening systems that take ophthalmic images, fundus photographs and sometimes optical coherence tomography scans, check that they are usable, run segmentation and grading models, and present structured findings to a clinician alongside the image with the regions that drove the result highlighted. Such a system can also generate a report and a suggested follow-up interval. What it does not do, in any defensible design, is replace the clinician making the decision.

The Challenge

Ungradable input silently scored as normal is the most dangerous failure mode in the whole pipeline, and it is the default behaviour of any model given an image degraded by media opacity, defocus, poor illumination or artefact. Model work is then dominated by generalisation rather than headline accuracy: performance drops across camera models, populations and imaging protocols, so a system validated only on its training sources will disappoint in clinic. Image identity is its own difficulty, because laterality, field position, pupil dilation status and capture device all change how a result should be read, and study, series and instance relationships are what let multiple fields and both eyes be assembled into one assessment. Software that grades pathology or drives referral is also a medical device in most markets, so silent retraining is not permissible and every result must be reproducible.

The Solution

Image ingestion and quality gating

Images arrive from cameras that speak DICOM, from vendor exports, or from a web upload.

• Handling DICOM properly means the ophthalmic photography object classes, retrieval and storage over DICOMweb, and consistent patient and study identity.
• Laterality, field position, pupil dilation status and capture device belong in the record.
• The single most valuable component is a gradability model that rejects degraded images before any pathology model runs.

Segmentation, grading and calibration

Anatomical segmentation of the optic disc, cup and vessel tree supports derived measures such as cup to disc ratio and vessel calibre. Lesion detection identifies microaneurysms, haemorrhages, exudates and neovascularisation, which feed grading against an established clinical scale rather than an invented one, so outputs align with what clinicians already use for referral decisions. External validation on data from sources unseen in training is mandatory. Probabilities are calibrated, operating points are chosen for sensitivity on referable disease rather than balanced accuracy, and the model can abstain and route to a person when confidence is low.

Application architecture

A workable stack is a Python web application handling identity, workflow and reporting, with inference isolated in worker processes on accelerated hardware so a slow model cannot block the request path. Images live in object storage with encryption and strict access control, while the database holds studies, results, model versions and audit trails. Health information must never reach application logs or error trackers. Integration with the wider record uses HL7 messaging or FHIR resources so results land where clinicians already work rather than in a separate silo.

Regulatory and clinical safety constraints

Device authorisation brings a documented software lifecycle, risk management, clinical evaluation and post-market surveillance. Model updates are not routine deploys: change control forms part of the authorisation. Every result stores the exact model version, thresholds and preprocessing that produced it. Saliency overlays help clinicians interrogate a result but are explanatory aids, not proof. Any product recommendation attached to the workflow stays clearly commercial and is never presented as treatment.

What This Delivers

Clinicians get structured findings next to the image, with the driving regions highlighted, a calibrated confidence and an explicit abstention where the system should not be trusted. Poor captures are caught while the patient is still in front of the camera rather than becoming a false reassurance. Results flow into the record clinicians already use, and every one of them can be reproduced later from the stored model version, thresholds and preprocessing. The practical priorities are to gate on image quality first, keep the clinician as decision maker, validate on representative populations, and monitor for drift as cameras and protocols change.

Technologies and Tools

• DICOM ophthalmic photography object classes with DICOMweb retrieval and storage
• Gradability, segmentation and lesion detection models with calibration and abstention
• Grading against established clinical scales, externally validated on unseen sources
• A Python web application for identity, workflow and reporting, with isolated inference workers on accelerated hardware
• Encrypted object storage, versioned result records and audit trails
• HL7 messaging and FHIR resources for record integration