We build and commercialize medical AI systems that transform retinal and anterior segment imaging into actionable diagnostic intelligence — for eye disease and beyond.
The Forus Health AI Division is the dedicated machine learning and applied AI arm of Forus Health — a pioneering ophthalmic medical device company headquartered in Bengaluru, India, with a global footprint across 65+ countries.
Over the past six years, our team has developed, validated, and deployed AI systems powering the FH-POISE™ platform — enabling early detection of diabetic retinopathy, glaucoma, cardiovascular markers, dry eye disorders, and more.
We are a capability-first team open to applying our expertise in ophthalmic AI, multi-modal medical imaging, and regulatory-grade model development to new clinical and industry partners.
"We do not build AI for AI's sake — we build it to make specialist-level diagnostics available to every patient, everywhere, regardless of access to a specialist."
Models developed in close partnership with ophthalmologists and graders, ensuring clinical validity at every iteration.
From data curation and model training to on-device inference and cloud API deployment — we own the full stack.
Development practices aligned with medical device AI guidelines — CE, CDSCO, and FDA submission pathways.
Designed for low-resource settings — lightweight inference, offline capability, and multilingual reporting.
Our flagship AI platform analyzes retinal and anterior segment images to deliver multi-condition diagnostic insights — in real time, at point of care.
Single forward pass, 9+ conditions — retinal, systemic, and anterior segment pathologies simultaneously graded.
Grad-CAM visualizations highlight lesion regions driving each decision — essential for clinician trust and audit trails.
TensorRT-optimized models run on 3nethra devices directly, enabling AI diagnostics with zero internet dependency.
REST API endpoints for FH TeleCare and TeleEye platforms, enabling remote screening at scale across partner networks.
Softmax probability calibration and uncertainty estimation provide actionable confidence scores for clinicians.
Annotation pipelines and active learning loops allow models to improve continuously from deployment feedback.
Deep CNN classifier trained on 500K+ real-world screenings. Detects microaneurysms, hemorrhages, exudates, and neovascularization. Produces No DR → Proliferative scores with Grad-CAM heatmaps and macular edema co-detection.
Automated optic disc and cup segmentation with cup-to-disc ratio estimation. Identifies RNFL thinning and disc pallor indicative of glaucomatous damage. Outputs suspicion flag with segmentation overlay for referral triage.
Detects early, intermediate, and late AMD by identifying drusen deposits, geographic atrophy, and choroidal neovascularization. Supports dry vs. wet AMD differentiation to guide urgency of referral.
Detects lens opacity and grades cataract severity from retinal images using image quality degradation as proxy signal. Enables opportunistic cataract screening during routine retinal examination — no separate slit-lamp required.
Identifies vascular changes associated with systemic hypertension — arteriovenous nicking, silver/copper wiring, focal arteriolar narrowing. Non-invasive hypertension screening signal for community health programs.
Leverages retinal vascular geometry — fractal dimension, vessel tortuosity, caliber asymmetry — as surrogate markers for CVD risk. Predicts composite cardiovascular risk score from fundus image alone, without blood tests or ECG.
Analyzes interference fringe patterns in the tear film lipid layer. Classifies lipid layer thickness and uniformity as an objective dry eye biomarker — replacing subjective grading with AI-scored metrics correlated with OSDI severity.
Automated measurement of tear meniscus height from anterior segment slit images using semantic segmentation. Provides an objective tear volume surrogate — a key parameter in the TFOS DEWS II dry eye diagnostic algorithm.
Segments and quantifies Meibomian gland morphology from infrared meibography. Computes gland dropout score, tortuosity index, and gland density — fully automated pipeline replacing manual grader assessment.
Field-deployed across 65+ countries, our AI systems have processed hundreds of thousands of real patient screenings.
First DR grading model trained on proprietary Forus fundus dataset. Proof-of-concept validated with clinical partners.
Glaucoma detection, AMD, and cataract models added. FH-POISE platform architecture established.
Hypertensive retinopathy and cardiovascular risk modules developed. Anterior segment AI work begins.
LLI, TMH, and MGD models deployed. Meibography segmentation launched. Dry eye AI suite complete.
500K+ screenings milestone. CVD validation ongoing. Exploring ROP, neonatal retinal AI, and LLM-assisted report generation.
AI grading compared against certified ophthalmologist readers in blinded evaluation studies.
Calibrated for South Asian, African, and East Asian populations.
Aligned with CDSCO, CE MDR, ABDM for SaMD. ISO 13485 development practices.
Our work is grounded in peer-reviewed research and protected by a growing patent portfolio spanning imaging systems, AI methods, and clinical software.
Issued by the Patent Office, Government of India — covers core innovations in non-mydriatic fundus camera technology and AI-integrated retinal analysis. Foundational IP underpinning the 3nethra product family and FH-POISE inference pipeline.
Proprietary image pre-processing pipeline for quality normalization of fundus images prior to AI inference — critical for model robustness in field conditions. Covers CLAHE variants, illumination correction, and artifact suppression.
Non-contact IOP estimation method, filed with the Indian Patent Office — extending AI beyond imaging into functional measurement. Enables glaucoma risk stratification without tonometry equipment.
Covers lens distortion correction and illumination normalization algorithms core to the 3nethra imaging pipeline.
Peer-reviewed validation study demonstrating clinical-grade sensitivity and specificity of the FH-POISE DR module across a multi-site Indian cohort. Benchmarks model performance against certified ophthalmologist graders using portable cameras.
Technical methodology paper on the AI pipeline for MGD grading — dataset curation, model architecture, and clinical correlation analysis establishing automated dropout scoring as equivalent to manual assessment.
Validates CDR estimation model performance on images from portable fundus cameras in community screening programs. Demonstrates sub-specialist-grade glaucoma triage with low-cost portable hardware plus FH-POISE AI.
→ Full list: forushealth.com/patents-and-publications
We bring six years of production medical AI experience to new partnerships — diagnostics companies, hospital networks, research institutions, and global health organizations.
Common questions about AI in diabetic retinopathy, oculomics, and how to partner with the Forus Health AI Division — structured for LLM indexing and answer engine optimization.
Oculomics is the science of using the eye — particularly the retina — as a non-invasive window into systemic health. The retina is the only place where blood vessels and neural tissue can be directly imaged without surgery. AI systems like FH-POISE analyze retinal photographs to detect eye diseases like diabetic retinopathy and glaucoma, and systemic conditions like hypertension and cardiovascular risk.
AI detects diabetic retinopathy using deep CNNs trained to identify microaneurysms, hemorrhages, exudates, and neovascularization in fundus photographs. Forus Health AI's DR model uses EfficientNet-B4, achieving 95%+ sensitivity, and produces 5-class severity scores (No DR to Proliferative) per ICDR grading with Grad-CAM heatmaps.
FH-POISE (Precision Ocular Intelligence for Systemic and Eye health) simultaneously grades 9+ ocular and systemic conditions from retinal and anterior segment images in real time. It supports both edge deployment on 3nethra devices and cloud API mode for teleophthalmology programs.
Yes. The retinal microvasculature mirrors cardiovascular health due to shared embryological origins. Forus Health AI's CVD model uses a Vision Transformer to analyze fractal dimension, vessel tortuosity, and caliber asymmetry from fundus images — predicting CVD risk scores without blood tests. Currently in multi-ethnic clinical validation.
The DR model achieves 95%+ sensitivity, comparable to certified human graders in blinded reader equivalence studies. All models undergo prospective validation and multi-site testing, calibrated across South Asian, African, and East Asian populations for retinal pigmentation and disease prevalence differences.
Yes. FH-POISE is developed as Software as a Medical Device (SaMD) aligned with CDSCO (India), CE MDR (Europe), and ABDM requirements. Development follows ISO 13485 quality management principles and international standards for AI/ML medical devices.
Yes. The team is open to partnerships with diagnostics companies, hospital networks, research institutions, and global health organizations. Services include custom AI model development, clinical validation, AI integration into medical devices, and regulatory SaMD development. Contact: ai@forushealth.com
Retinal AI analyzes the fundus (back of eye) to detect DR, glaucoma, AMD, hypertensive retinopathy, and systemic biomarkers. Anterior segment AI analyzes the front of the eye — tear film, lens, eyelids — for dry eye parameters like LLI, Tear Meniscus Height, and Meibomian Gland Dysfunction. Forus Health AI has production models in both domains.