Related Experiment Videos
Computerized detection of clustered microcalcifications in digital mammograms using a shift-invariant artificial
1Kurt Rossmann Laboratories for Radiologic Image Research, Department of Radiology, University of Chicago, Illinois 60637.
Medical Physics
|April 1, 1994
Summary
A novel shift-invariant neural network effectively reduces false positives in digital mammography computer-aided diagnosis (CAD). This AI approach enhances detection accuracy for clustered microcalcifications, improving diagnostic reliability.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Biomedical Engineering
Background:
- Computer-aided diagnosis (CAD) schemes are crucial for detecting clustered microcalcifications in digital mammograms.
- False-positive detections from CAD systems can reduce diagnostic efficiency.
Purpose of the Study:
- To apply a shift-invariant neural network (SI-NN) to reduce false-positive findings from a CAD scheme.
- To improve the accuracy of clustered microcalcification detection in mammograms.
Main Methods:
- A shift-invariant neural network (SI-NN), a multilayer back-propagation network with local, shift-invariant interconnections, was utilized.
- The SI-NN was trained to detect individual microcalcifications within regions of interest (ROIs) identified by the CAD scheme.
- Performance was evaluated using jackknife resampling and Receiver Operating Characteristic (ROC) analysis on 168 ROIs from 34 mammograms.
Main Results:
- The SI-NN achieved an average area under the ROC curve (Az) of 0.91.
- Approximately 55% of false-positive ROIs were eliminated without compromising true-positive detections.
- Performance significantly surpassed that of a conventional three-layer, feed-forward neural network.
Conclusions:
- Shift-invariant neural networks offer a superior method for reducing false positives in mammogram CAD systems.
- This approach enhances the reliability and efficiency of detecting clustered microcalcifications.
- The location-independent nature of SI-NNs contributes to their improved performance in medical image analysis.