Mayo Clinic Diverticulosis Detection
AI-Powered Medical Imaging for Clinical Diagnosis
Role
UX Research and Design, Co-developer
Problem
Diverticulosis affects 60% of Americans over 58, with significant miss-rates during colonoscopy procedures. Doctors can miss small diverticular bulges and even large polyps during routine examinations, leading to delayed diagnosis and treatment of this devastating condition.
Solution
Created custom machine learning and AI algorithms that power a real-time computer-aided detection system streaming video from endoscopes during colonoscopies. These proprietary algorithms I developed are now actively used at Mayo Clinic to highlight diverticular disease characteristics, automatically slowing video to half frame rate and saving pictures for confirmation when potential matches are detected.
System Overview
Computer-Aided Detection System
Built a system that streams video from endoscopes and uses AI and computer vision algorithms to highlight possible characteristics of diverticular disease, presenting them to doctors for verification.
AI Algorithms in Clinical Use
Developed proprietary machine learning algorithms that are now actively deployed at Mayo Clinic. These custom AI detection models I created continue to assist physicians in real-time during colonoscopy procedures, representing a lasting contribution to clinical AI and computer-assisted medical diagnosis.
Clinical Integration
Worked directly with doctors, observing colonoscopy procedures and understanding useful features needed to assist during examinations. Translated clinical requirements into a working prototype designed specifically for Mayo Clinic's Windows environment.
Sample Video: Doctor's View with Side-by-Side Detection
Technical Challenges Solved
Strong Light Reflection
Reflection from camera lighting caused noisy edges in binary images, especially in water-filled diverticulars.
Limited Texture Information
Monotone color and minimal texture information inside the colon made feature extraction challenging.
Shape Variance
Large variance in diverticular shape and appearance from different viewing angles and between patients.
Scale and Rotation Variance
Distance and camera rotation during procedures caused significant variance in scale and orientation.
System Architecture
Frame Capturer
Generates JPEG frames from MPEG video with configurable frame rates
AI Detection Engine
Machine learning algorithms for pattern recognition and classification
Edge Detection
Canny edge detector and Laplacian of Gaussian filters for contour extraction
Ellipse Fitting Model
Identifies elliptical shapes characteristic of diverticular disease
System Flow Overview
Processing Pipeline
Detection Methodology
Edge Detection & Filtering
Advanced algorithms for identifying diverticular contours
Pattern Recognition
Machine learning classification of detected features
Ellipse Fitting
Geometric modeling of diverticular shapes
Feature Extraction & Classification
I designed and implemented proprietary machine learning algorithms that employ sophisticated pattern recognition to distinguish between diverticular disease and normal colon features. Developed custom classification models including Naive Bayesian, Decision Trees, and Support Vector Machines, optimizing them for clinical deployment at Mayo Clinic.
Pattern Templates
Grayscale and HSV color space analysis for feature classification
Temporal Tracking
Sequential frame analysis for enhanced detection accuracy
Clinical Impact & Results
Americans Over 58
Affected by diverticulosis
Detection Rate
Sensitivity & specificity
Processing
During live procedures
My AI Algorithms Now Used at Mayo Clinic
I created the core machine learning and AI algorithms that power Mayo Clinic's diverticulosis detection system. These proprietary algorithms I developed continue to assist physicians during live colonoscopy procedures, representing a lasting contribution to clinical AI. The system pioneered real-time AI detection in medical procedures, helping reduce miss-rates for diverticular disease detection while maintaining seamless clinical workflow integration.