Deep Conformal Supervision: Leveraging Intermediate Features for Robust Uncertainty Quantification.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning
• Medical Imaging

Background:

• Trustworthy artificial intelligence (AI) is essential for clinical applications.
• Uncertainty quantification (UQ) is a key component of trustworthy AI.
• Conformal prediction is a robust UQ framework gaining traction in AI.

Purpose of the Study:

• To introduce Deep Conformal Supervision (DCS) for improved non-conformity score calculation in conformal prediction.
• To enhance the trustworthiness of AI models in clinical settings through better UQ.
• To evaluate DCS performance on medical image classification tasks.

Main Methods:

• Leveraging intermediate outputs from deep supervision for non-conformity scores.
• Employing weighted averaging based on inverse mean calibration error across stages.
• Benchmarking on pneumonia chest radiography and intracranial hemorrhage datasets.

Main Results:

• DCS achieved significantly lower mean coverage errors compared to baseline methods on both datasets (p < 0.001).
• Observed improvements were particularly pronounced in scenarios with smaller datasets.
• The method demonstrated enhanced performance when smaller acceptable error levels were considered.

Conclusions:

• Deep Conformal Supervision offers a significant advancement in UQ for healthcare AI.
• The method is particularly valuable for improving AI reliability in data-scarce medical imaging scenarios.
• DCS contributes to developing more trustworthy AI systems for clinical decision support.