Reliable deep-learning-based phase imaging with uncertainty quantification.

Area of Science:

• Biomedical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Deep learning (DL) shows promise for revolutionizing biomedical imaging.
• A key challenge is the lack of reliability assessment for DL predictions, with errors often discovered post-hoc.
• Existing methods struggle to quantify uncertainty in DL models.

Purpose of the Study:

• To develop a Bayesian convolutional neural network (BNN) framework for quantifying uncertainty in DL predictions for biomedical imaging.
• To demonstrate that BNN-predicted uncertainty maps can serve as reliable indicators of DL model and measurement errors.
• To enable quantitative credibility assessment of DL-generated images.

Main Methods:

• Proposed a novel Bayesian convolutional neural network (BNN) framework.
• Developed uncertainty maps to estimate prediction errors and characterize imperfections (noise, model error, data limitations).
• Applied the BNN to large space-bandwidth product phase imaging using a physics-guided coded illumination scheme.

Main Results:

• BNN-predicted uncertainty maps accurately estimate true errors from the model and measurements.
• Uncertainty maps revealed imperfections like noise, model error, and out-of-distribution data.
• Demonstrated gigapixel phase image prediction from minimal measurements with quantitative credibility assessment.
• Identified rare biological phenomena in low-certainty regions.

Conclusions:

• The BNN framework provides a reliable method for assessing DL prediction confidence and data/model quality in biomedical imaging.
• This uncertainty quantification is crucial for trustworthy AI in healthcare.
• The framework is broadly applicable to various DL-based biomedical imaging techniques.