Medtronic

GI Genius

FDA 510(k) Cleared AI for Polyp Detection

Problem
26% of polyps missed during standard colonoscopy; regulatory pathway unclear for AI-assisted devices
Solution
Guided technical strategy and regulatory documentation for first AI colonoscopy device to achieve FDA clearance
Medtronic GI Genius Platform
FDA510(k) Cleared
99.7%Sensitivity
14%More polyps found
30fpsReal-time inference

26% of polyps are missed. Every miss is a potential cancer.

Colorectal cancer is the second leading cause of cancer death in the US. Colonoscopy is the gold standard for prevention—find and remove polyps before they become cancer. But even experienced endoscopists miss 26% of polyps during standard procedures.

The technology existed to help: deep learning models could detect polyps in real-time video with high accuracy. But the path to clinical deployment was unclear. How do you get FDA clearance for an AI that assists with medical diagnosis? What evidence do you need? What documentation?

I advised on the technical and regulatory strategy. Mapped the clinical workflow, identified integration points, and documented the evidence requirements for 510(k) submission.

26%polyp miss rate
15Mprocedures/year
#2cancer killer
Colonoscopies Performed: 15.0M (100%)15.0MColonoscopies Performed100%Polyps Present: 6.0M (40.0%)6.0MPolyps Present40.0%60.0%Polyps Detected: 4.2M (70.0%)4.2MPolyps Detected70.0%30.0%Adenomas Found: 2.1M (50.0%)2.1MAdenomas Found50.0%50.0%Cancer Prevented: 840.0K (40.0%)840.0KCancer Prevented40.0%60.0%
Total Visitors:15.0M
Conversions:840.0K
Overall Rate:5.60%

30 frames per second. Zero disruption to workflow.

The AI module connects between the endoscope and the display. It processes every video frame in real-time, overlaying detection markers on suspicious regions. The physician sees the same view they always have—plus AI assistance.

Architecture constraints: must work with existing equipment, must not add latency visible to the physician, must integrate with EHR for documentation, must maintain complete audit trail for regulatory compliance.

Latency: <33ms end-to-end (invisible to user)
Integration: Works with all major endoscope brands
Documentation: Auto-generates procedure findings
Loading diagram...

99.7% sensitivity. Validated in multi-center trials.

The model was trained on 13 million+ frames from over 5,000 procedures. Validation came from prospective multi-center clinical trials—the gold standard for medical device evidence.

Key to FDA clearance: comprehensive documentation of model performance, failure modes, and limitations. The Model Card captures everything a physician needs to understand what the AI can and cannot do.

99.7%sensitivity
98.5%specificity
33msinference time
Polyp Detection Performance
GI Genius Polyp Detection v2.1
Area Under Curve (AUC)
0.987
Excellent
0%0%20%20%40%40%60%60%80%80%100%100%Random (AUC = 0.5)Clinical ThresholdSens: 99.7%Spec: 98.5%False Positive Rate (1 - Specificity)True Positive Rate (Sensitivity)
Legend
GI Genius Polyp Detection v2.1
Human Expert (Mean) (AUC: 0.890)
Random classifier
Operating Point
99.7%
Sensitivity
98.5%
Specificity

AI Model Card — Regulatory Documentation

Comprehensive model documentation required for FDA 510(k) submission.

Intended Use
Aid to endoscopists in detecting colorectal polyps during colonoscopy procedures in real-time
Users: Gastroenterologists, Endoscopists, Colorectal Surgeons
Out of Scope
Diagnostic determination without physician review
Pediatric colonoscopy procedures
Non-colonoscopy GI procedures
Training Data
Dataset Size
13M+ video frames
Date Range
2018 - 2023
Features
Raw video frames at 30fps
Train/Test Split
70/15/15
Deployment
Embedded device (no network)
Latency
33ms (30fps)
Throughput
Real-time video processing
Limitations
Requires adequate bowel preparation (Boston ≥6)
Performance may vary with withdrawal speed (<6min)
Does not replace physician clinical judgment
Not validated for pediatric patients

From clearance to 500+ sites in 12 months.

FDA clearance was the beginning, not the end. Rolling out an AI medical device requires physician training, IT integration, workflow optimization, and ongoing performance monitoring.

Built deployment playbooks covering installation, training, and support. Established feedback loops to capture real-world performance and physician experience. Created dashboards tracking adoption and clinical outcomes.

Before AI

  • ✗ 26% polyp miss rate
  • ✗ Physician-dependent quality
  • ✗ No real-time assistance

With GI Genius

  • ✓ 14% improvement in ADR
  • ✓ Consistent AI assistance
  • ✓ Real-time detection overlay
Sites Deployed
4066.7%
500
Min: 12Max: 500
Procedures/Month
18122.2%
82K
Min: 450Max: 82K
ADR Improvement
75%
+14%
Min: +8%Max: +14%
Physician NPS
32.3%
82
Min: 62Max: 82

The Result

GI Genius received FDA 510(k) clearance in April 2021—the first AI system cleared for real-time polyp detection during colonoscopy. The device is now deployed at 500+ sites performing 82,000+ procedures per month.

Clinical studies show a 14% improvement in adenoma detection rate (ADR) when using GI Genius. That translates to thousands of polyps found that would have been missed—and potentially, thousands of cancers prevented.

No AIFDA ClearedRegulatory Status
26% miss+14% ADRDetection Rate
0 sites500+ sitesDeployment Scale
Clinical Outcomes
510(k)
FDA Cleared
14%
↑ ADR Detection
1000+
Installations

Technology Stack

Core technologies powering the FDA-cleared AI medical device.

AI/ML
PyTorch
Deep learning
TensorFlow
Model serving
Embedded
Docker
Edge deployment
Python
ML development
Regulatory
FDA
510(k) pathway
HIPAA
Data protection