Clinical AI — Model Library

Five models.
Four specialties.

Every Thornex AI model is independently validated on held-out clinical datasets before deployment. We publish our accuracy figures, our confidence intervals, and our failure modes — because you deserve to know exactly what you are working with.

HIPAA-compliant Zero data retention Sub-second inference
5Active models
99%Peak accuracy
<1sInference time
99.1%Peak accuracyChest X-Ray module
0.8sFastest inferencePer analysis request
5Clinical modelsAcross 4 specialties
0 KBData retainedHIPAA zero-retention
Clinical Models

Every module, explained

Five specialised AI models, each validated independently and built for real clinical workflows — not demos.

DermatologyProductionDERM v1.2

Dermatological AI

Lesion Classification

Dermoscopic lesion analysis using a three-stage ensemble architecture. Classifies melanoma, basal cell carcinoma, squamous cell carcinoma, and 14 additional dermatological conditions with clinical-grade precision.

Trained on 2.4M verified dermoscopic images
Malignancy risk stratification with confidence scores
Binary mask and topological segmentation overlay
HIPAA-compliant zero-retention pipeline
U-Net EnsemblePyTorchCUDA 11.8ONNX Export
Open Module
Dermatological AI
Accuracy
97.8%
Inference Speed
0.9s
Tech Stack
U-Net EnsemblePyTorchCUDA 11.8ONNX Export
DERM v1.2Production
OphthalmologyProductionODIR v2.5

Ocular Diagnostic AI

Retinal Fundus Analysis

Full-spectrum fundus image classification for 8 ocular pathologies including diabetic retinopathy, glaucoma, macular degeneration, and hypertensive retinopathy. Trained on the ODIR-5K clinical benchmark.

Dual-ring retinal feature extraction pipeline
Optic disc and macula boundary detection
Confidence-ranked differential diagnosis output
Compatible with standard fundus camera DICOM formats
ViT-LargeTensorRTONNXDICOM Ready
Open Module
Ocular Diagnostic AI
Retinal Zones Analysed
Macula
Optic Disc
Periphery
Accuracy
98.5%
Inference Speed
1.1s
RadiologyProductionPULMO v1.0

Pulmonary Radiological AI

Chest X-Ray Triage

AI-assisted chest radiograph analysis for pneumonia, pleural effusion, cardiomegaly, infiltrates, and pulmonary nodule patterns. Engineered for high-throughput emergency department workflows.

Trained on NIH ChestX-ray14 and proprietary clinical dataset
Anomaly localisation with precision bounding box overlay
Three-tier severity scoring: mild, moderate, critical
Sub-second critical finding flag for emergency triage
ResNet-152PyTorchCUDA 11.8HL7 FHIR
Open Module
Pulmonary Radiological AI
Finding Severity Distribution
Clear55%
Moderate30%
Critical15%
Peak Accuracy
99.1%
Clinical NutritionProductionNUTRI v1.0

Metabolic Planning AI

AI Diet Planner

Evidence-based clinical nutrition planning from patient body metrics, lifestyle parameters, and therapeutic goals. Generates personalised daily meal plans with complete macronutrient and hydration targets.

BMR, TDEE, and BMI calculation pipeline
Calorie-scaled meal composition per dietary goal
Hydration targets adjusted for activity level
Supports weight loss, maintenance, and therapeutic gain
LLM-BackedREST APIJSON OutputEHR Export
Open Module
Metabolic Planning AI
Protein30%
Carbs45%
Fat25%
Inference Speed
1.4s
NUTRI v1.0Production
RadiologyBetaSEG v2.1

Pixel-Level Structural AI

Radiological Segmentation

High-precision U-Net architecture for delineating structural anomalies in CT and MRI volumes. Produces pixel-perfect organ and lesion boundary maps for surgical planning and oncological staging.

Volumetric segmentation across 12 organ classes
Sub-voxel precision lesion boundary delineation
Real-time 3D rendering pipeline
Validated on 180,000+ annotated clinical scans
U-Net 3DPyTorchCUDA 11.8NIfTI / DICOM
Open Module
Pixel-Level Structural AI
Accuracy
95.1%
Inference Speed
1.6s
Tech Stack
U-Net 3DPyTorchCUDA 11.8NIfTI / DICOM
SEG v2.1Beta

Ready to run your
first analysis?

Create a physician account and access all five clinical modules immediately — no setup, no waiting.

Open Workspace