Comparative evaluation of uncertainty estimation and decomposition methods on liver segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) require reliable uncertainty estimation for high-risk applications like medical decision support.
• Predictive uncertainty in DNNs can be decomposed into aleatoric (data) and epistemic (model) components.

Purpose of the Study:

• To systematically review and compare methods for decomposing predictive uncertainty in Bayesian neural networks (BNNs).
• To evaluate mutual information decomposition against explicit modeling using a loss-attenuating neuron for uncertainty decomposition.

Main Methods:

• Experiments conducted on liver segmentation in CT scans.
• Qualitative evaluation of uncertainty maps and quantitative comparison of decomposed uncertainties.
• Analysis across varying training set sizes, label noise, and distribution shifts.

Main Results:

• Mutual information decomposition robustly produced meaningful aleatoric and epistemic uncertainty estimates.
• The loss-attenuating neuron exhibited noisier convergence and did not significantly improve segmentation or calibration.
• The heteroscedastic neuron showed marginal benefits in uncertainty quality but no significant performance gains.

Conclusions:

• Mutual information decomposition is a simple, mathematically sound method for obtaining reliable uncertainty estimates.
• Extending BNNs with loss-attenuating neurons offered no significant improvement in segmentation performance or calibration for this task.
• The study highlights the efficacy of mutual information decomposition for uncertainty quantification in medical imaging AI.