Uncertainty CNNs: A path to enhanced medical image classification performance.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Computational Biology

Background:

• Automated medical image analysis is crucial for early disease detection.
• Deep learning (DL) methods, particularly convolutional neural networks (CNNs), are widely used.
• Uncertainty quantification (UQ) is underdeveloped in medical imaging despite its importance for decision-making.

Purpose of the Study:

• To introduce a low-complexity, uncertainty-based CNN architecture for medical image classification.
• To quantify predictive uncertainty using a novel test-set augmentation technique.
• To demonstrate that UQ can improve classification performance in medical imaging tasks.

Main Methods:

• Developed a CNN architecture incorporating uncertainty quantification.
• Employed test-set augmentation to generate image surrogates and empirical distributions.
• Calculated mean estimates and credible intervals for uncertainty assessment.
• Evaluated the model on brain MRI, lung CT, and cardiac MRI datasets.

Main Results:

• The proposed method quantifies predictive uncertainty (aleatoric uncertainty).
• Test-set augmentation significantly improved classification performance across all datasets.
• The model demonstrated robustness against overfitting due to its low-complexity design.
• Achieved enhanced diagnostic accuracy for tumor and heart failure detection.

Conclusions:

• Test-set augmentation is a viable method for improving DL model performance in medical imaging.
• The developed uncertainty-based CNN provides reliable UQ and enhances diagnostic capabilities.
• The model's low complexity and re-trainability make it suitable for diverse clinical applications.