Deep Learning · CNN · Vision Transformers · GradCAM

Medical Imaging AI That Saves Lives

Radth builds lightweight CNN and Vision Transformer models that classify, triage, and risk-stratify medical images — from GI endoscopy to radiology, pathology, and beyond. Deployable on clinical hardware. Explainable by design.

<8M
Parameters
0
Missed High-Risk
98%+
AUC-ROC
HIPAA
Compliant
Medical Imaging AI Solutions

CNN Models for Every Clinical Modality

Lightweight, explainable, and clinically validated deep learning models deployable on standard hospital hardware — no GPU cluster required.

🔬Flagship

GI Endoscopy AI

Four-class risk stratification from gastrointestinal endoscopy images — Normal, Inflammatory, Pre-malignant, High-Risk — aligned with ACG/ESGE clinical guidelines.

DenseNet-121EfficientNet-B0DeiT-TinyAEL LossGradCAM
0.83
Macro F1
0.98
AUC-ROC
0
Missed HR
🫁Available

Chest Radiology AI

CNN-based analysis of chest X-rays and CT scans for pneumonia, nodule detection, pleural effusion, and COVID-19 screening with radiologist-level accuracy.

ResNet-50EfficientNetGrad-CAM++Ensemble
96%
Sensitivity
94%
Specificity
0.97
AUC
🧫Available

Digital Pathology AI

Whole-slide image analysis for cancer grading, cell segmentation, and Ki-67 proliferation index scoring — automated pathology at scale.

ViT-BU-NetAttention MILPatch CNN
91%
Gleason Acc.
0.89
Cell mAP
0.96
AUC
🧬Available

Dermatology AI

Skin lesion classification across 7 diagnostic categories — melanoma, basal cell carcinoma, keratosis, and more — with dermoscopy image support.

EfficientNet-B4Vision ViTTTAMC Dropout
0.94
Melanoma AUC
0.87
Macro F1
92%
Sensitivity
👁️Available

Ophthalmology AI

Retinal image analysis for diabetic retinopathy grading (0–4 severity), glaucoma risk screening, and age-related macular degeneration detection.

InceptionV3ResNetCAMBootstrap CI
0.86
DR Kappa
0.95
Glaucoma AUC
93%
AMD Acc.
⚙️Bespoke

Custom CNN Pipeline

End-to-end custom model development for any medical imaging modality — from dataset curation and annotation to training, validation, and clinical deployment.

Any ArchitectureCustom LossActive LearningMLOps
8 wks
Delivery
12 mo
Support
Pathway
FDA

Need a model for a modality not listed? We build custom pipelines for any imaging type.

Discuss Your Imaging Problem
Our Technology

Research-Grade AI, Clinical-Ready

We translate published deep learning research into production-grade medical imaging software — with the explainability and safety guarantees that clinical deployment demands.

🧠

Clinically Aligned Loss

We encode clinical risk asymmetry directly into the training objective — not as post-hoc thresholds.

🔍

GradCAM Explainability

Every prediction comes with saliency maps showing exactly what the model attended to.

📊

Uncertainty Quantification

Monte Carlo Dropout flags low-confidence cases for mandatory clinician review — structurally safe.

Lightweight by Design

All models <10M parameters. Inference on CPU in <200ms. No GPU cluster required for deployment.

Our Development Pipeline

From raw clinical images to a validated, deployed model — we own the full stack. Every project follows our 5-stage pipeline built around clinical safety requirements.

01

Dataset & Annotation

Clinical dataset curation, class mapping to guidelines, and HIPAA-compliant annotation workflows.

02

Model Architecture

CNN + Vision Transformer selection optimised for your modality. Lightweight (<10M params) for clinical deployment.

03

Custom Loss + Training

Asymmetric loss functions (AEL-style) encode clinical cost asymmetry. Monte Carlo Dropout for uncertainty.

04

Explainability & Validation

GradCAM heatmaps, bootstrap CIs, and clinical validation studies — results your clinicians can trust.

05

Clinical Deployment

Docker/ONNX packaging, DICOM integration, HL7 FHIR support — runs on your existing hospital infrastructure.

Architecture Comparison

DenseNet-121CNN
7.0M params
Macro F10.83
AUC-ROC0.975
EfficientNet-B0CNN
5.3M params
Macro F10.83
AUC-ROC0.973
DeiT-TinyTransformer
5.9M params
Macro F10.81
AUC-ROC0.969

Results on HyperKvasir GI endoscopy benchmark

6+
Imaging Modalities
<8M
Params Per Model
0.98+
Mean AUC-ROC
0
Missed High-Risk
Get In Touch

Ready to Deploy Your AI Agent?

Tell us about your business and we'll recommend the right AI agent. Most clients are live within 2 weeks.

Contact Radth AI

Location

New York, NY

Follow Us

Typically replies within 2 hours

Mon–Fri, 9am–6pm EST

Send us a message