Consensus versus disagreement in imaging research: a case study using the LIDC database.

Area of Science:

• Medical Imaging Analysis
• Machine Learning in Healthcare
• Diagnostic Systems

Background:

• Traditional computer-aided diagnosis (CAD) evaluation relies on single expert labels versus single CAD outputs.
• Intra-reader variability in medical image interpretation presents a challenge for diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a novel CAD system utilizing a Belief Decision Tree algorithm.
• To enable the system to learn from probabilistic input, accounting for intra-reader variability.
• To provide probabilistic output for enhanced diagnostic decision-making.

Main Methods:

• Implementation of a Belief Decision Tree classification algorithm for probabilistic learning and output.
• Comparison of the probabilistic CAD approach with a traditional decision tree method.
• Evaluation using standard accuracy and probabilistic area under the distance-threshold curve (AuC(dt)) metrics.
• Validation using cross-validation on training and validation subsets, and comparison with Lung Image Database Consortium data.

Main Results:

• The probabilistic CAD system demonstrated significant performance improvements over the traditional approach.
• Accuracy boosts of 28.26% (training) and 20.64% (validation) were observed.
• Area under the distance-threshold curve (AuC(dt)) boosts of 30.28% (training) and 23.21% (validation) were noted.
• The system's errors were associated with low confidence, and it agreed with benign diagnoses more often than radiologists.

Conclusions:

• The probabilistic CAD system offers superior performance compared to traditional methods, particularly in handling diagnostic uncertainty.
• The system's ability to provide probabilistic output and learn from variability enhances diagnostic reliability.
• This approach has the potential to reduce false positives in medical image analysis and assist radiologists.