Uncertainty quantification via localized gradients for deep learning-based medical image assessments

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are crucial for medical image assessment but require reliable measures beyond accuracy.
• Existing uncertainty quantification (UQ) methods often lack post hoc compatibility and alter model outputs.
• There is a need for UQ methods that can be applied to already validated models without retraining.

Purpose of the Study:

• To introduce and validate a novel post hoc UQ method, Local Gradients UQ, for deep learning-based medical image analysis.
• To demonstrate the utility of Local Gradients UQ in enhancing the reliability of metastatic disease delineation.
• To compare the performance of Local Gradients UQ against non-gradient UQ measures.

Main Methods:

• Developed Local Gradients UQ, a post hoc method leveraging a model's localized gradient space to assess parameter sensitivities.
• Compared Local Gradients UQ with non-gradient UQ methods based on model probability outputs.
• Evaluated UQ methods across four experiments: image quality degradation, high- vs. low-quality image comparison, false positive filtering, and correspondence with physician assessments.

Main Results:

• Local Gradients UQ showed significantly higher sensitivity to image quality degradation compared to non-gradient methods.
• The method demonstrated superior performance in distinguishing between high- and low-quality clinical images (p<0.05).
• Local Gradients UQ improved false positive filtering (ROC AUC +20.1%) and correspondence with physician likelihood assessments (ROC AUC +16.2%).

Conclusions:

• Local Gradients UQ is a novel, effective gradient-based UQ method for deep learning in medical imaging.
• The method enhances user trust by providing reliable indicators of model performance.
• Local Gradients UQ offers a valuable tool for deploying deep learning models in clinical settings.