Related Experiment Videos
Computer-assisted detection of colonic polyps with CT colonography using neural networks and binary classification
Anna K Jerebko1, Ronald M Summers, James D Malley
1Department of Radiology, National Institutes of Health, 10 Center Drive, Bethesda, Maryland 20892-1182, USA.
Medical Physics
|February 1, 2003
Summary
Neural networks achieved 90% sensitivity and 95% specificity in detecting colonic polyps via CT colonography, outperforming recursive binary trees and offering improved computer-aided detection for this challenging screening method.
Area of Science:
- Medical Imaging
- Computer-Aided Diagnosis
- Gastroenterology
Background:
- Detecting colonic polyps in CT colonography is challenging due to complex polyp shapes and normal colon surface variations.
- Existing computer-aided detection (CAD) methods show feasibility but require enhanced classifiers for improved specificity.
Purpose of the Study:
- To compare the classification performance of neural networks and recursive binary trees for colonic polyp detection.
- To identify a superior computational approach for improving the accuracy of polyp identification in CT colonography.
Main Methods:
- Surface geometry information was extracted from 3D colon reconstructions.
- A filter utilizing region density, Gaussian and average curvature, and sphericity identified candidate polyps.
- Neural networks and recursive binary trees were applied to classify candidate sites.
Main Results:
- A dataset of 39 polyps (3-25 mm) was analyzed using tenfold cross-validation.
- The backpropagation neural network with a single hidden layer, trained via the Levenberg-Marquardt algorithm, demonstrated superior performance.
- This neural network achieved 90% sensitivity and 95% specificity, with 16 false positives per study.
Conclusions:
- Neural networks, specifically the backpropagation model, offer a more effective classification approach for colonic polyps in CT colonography.
- The developed method shows significant potential for enhancing the specificity of computer-aided polyp detection, addressing a key limitation in current systems.