Mitigating Diagnostic Errors in Lung Cancer Classification: A Multi-Eyes Principle to Uncertainty Quantification.

Area of Science:

• Radiology
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Diagnostic errors and cognitive biases challenge accurate lung cancer diagnosis in radiology.
• Experienced radiologists often overlook perceptual and interpretive errors, as well as biases like anchoring and premature closure.

Purpose of the Study:

• To introduce the Multi-Eyes principle, leveraging multiple deep learning models to mitigate bias and enhance diagnostic accuracy.
• To address the limitations of single-model systems in computer-aided diagnosis.

Main Methods:

• Employed the Multi-Eyes principle, inspired by the Four-Eyes principle, using diverse 3D and 2D deep learning architectures.
• Integrated three uncertainty quantification techniques: Monte Carlo Dropout, Deep Ensemble, and Ensemble Monte Carlo Dropout.
• Utilized entropy for uncertainty measurement, averaged across models, followed by ensemble averaging of predictions on the LIDC-IDRI dataset.

Main Results:

• Statistical analysis showed that increased model numbers led to more peaked and left-skewed uncertainty distributions for incorrect predictions, indicating consensus.
• The Multi-Eyes principle demonstrated improvements in accuracy and F1 scores compared to single-model approaches.
• Effectively addressed overconfidence issues inherent in single deep learning models for lung cancer classification.

Conclusions:

• The Multi-Eyes principle shows significant potential for enhancing diagnostic performance in computer-aided diagnostic systems.
• This methodology offers a robust strategy to reduce bias and improve reliability in radiological diagnoses.
• Future work could explore advanced uncertainty quantification methods and feedback mechanisms.