Inference time correction based on confidence and uncertainty for improved deep-learning model performance and...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• Class imbalance in training data and limited explainability are major challenges for AI in medical image analysis.
• Existing methods often require a trade-off between model performance and explainability.

Purpose of the Study:

• To introduce a novel post-processing algorithm, CEbUTAl, to enhance both performance and explainability of AI models in medical imaging.
• To address class imbalance and improve trustworthiness of AI models in clinical settings.

Main Methods:

• Developed the Confidence and Entropy-based Uncertainty Thresholding Algorithm (CEbUTAl) as a model-agnostic, task-agnostic post-processing technique.
• Applied CEbUTAl to five medical imaging tasks, including intracranial hemorrhage detection and breast cancer detection, across various deep learning architectures and loss functions.

Main Results:

• CEbUTAl improved classification accuracy by approximately 5% and increased sensitivity across multiple tasks and models.
• Outperformed state-of-the-art methods in addressing class imbalance and quantifying uncertainty.
• Demonstrated that enhanced explainability does not necessitate a compromise in AI model performance.

Conclusions:

• CEbUTAl offers a generalizable approach to mitigate biases from class imbalance and improve AI explainability in medical imaging.
• The algorithm enhances the utility and trustworthiness of AI models for clinical practice.