Analysis and minimization of overtraining effect in rule-based classifiers for computer-aided diagnosis

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Radiology Decision Support

Background:

• Computer-aided diagnostic (CAD) schemes assist radiologists in lesion detection.
• Classifiers are crucial for high detection rates and low false positives in CAD.
• Traditional rule-based classifiers suffer from manual design, poor reproducibility, and overtraining.

Purpose of the Study:

• To develop a fully automated rule-based classifier.
• To introduce an optimal method for cutoff threshold selection.
• To minimize the overtraining effect in rule-based CAD classifiers.

Main Methods:

• Developed a fully automated rule-based classifier.
• Implemented an "optimal" method for cutoff threshold selection.
• Validated using Monte Carlo simulations and a lung nodule CT dataset.

Main Results:

• The automated threshold selection method completely eliminated overtraining during threshold selection.
• The overall overtraining effect in the constructed rule-based classifier was minimized.
• Demonstrated improved practicality and credibility of rule-based classifiers.

Conclusions:

• The proposed automated threshold selection method is highly effective in minimizing overtraining.
• This method significantly enhances the construction of reliable automated rule-based classifiers.
• The approach offers a valuable tool for improving CAD systems in medical imaging.