Determining Thresholds for Computer-Aided Detection for Silicosis-An Analytic Approach

Area of Science:

• Medical Imaging
• Occupational Health
• Pulmonary Medicine

Background:

• Computer-aided detection (CAD) is increasingly used for pulmonary tuberculosis (TB) screening via chest X-ray (CXR).
• CAD for silicosis, a significant TB risk factor, is less developed and requires expert validation.
• Selecting appropriate thresholds for CAD classification is critical for disease detection accuracy.

Purpose of the Study:

• To present an analytic approach for selecting thresholds in computer-aided detection (CAD) systems for silicosis.
• To explore the impact of different criteria, definitions, and reader variability on threshold selection.
• To introduce a novel two-threshold technique for improved classification of uncertain cases.

Main Methods:

• Compared Youden Index and minimum 90% sensitivity criteria using receiver operating curve data from a published study.
• Analyzed the influence of silicosis definition, expert reader, and disease prevalence on threshold choice and interpretation.
• Developed a novel method using two CAD thresholds to differentiate positive, negative, and uncertain cases.

Main Results:

• Thresholds varied significantly based on the chosen criterion, silicosis definition, and expert reader.
• Disease prevalence substantially impacted positive predictive value (PPV) and false positive rates.
• The study outlines implications for screening active miners, former miners, and compensation claims.

Conclusions:

• Users must define the use case, sensitivity/specificity preferences, and silicosis definition when applying CAD.
• System developers should account for and transparently report inter-reader variability in validation.
• A two-threshold model offers potential benefits for high-volume screening with high referral costs.