Metric-Guided Conformal Bounds for Probabilistic Image Reconstruction.

Area of Science:

• Medical Imaging
• Deep Learning
• Conformal Prediction

Background:

• Deep learning reconstruction algorithms create realistic medical scans from limited data.
• These algorithms can introduce inaccuracies, hindering statistically guaranteed claims about patient states.
• Existing methods struggle to provide reliable assessments from reconstructed images.

Purpose of the Study:

• To develop a framework for computing provably valid prediction bounds for probabilistic black-box image reconstruction.
• To ensure statistically guaranteed claims about the true state of a subject from reconstructed scans.
• To improve the reliability and interpretability of deep learning-based medical image analysis.

Main Methods:

• Representing reconstructed scans using a derived clinical metric of interest.
• Calibrating prediction bounds on the ground truth metric with conformal prediction (CP).
• Utilizing a prior calibration dataset for accurate bound estimation.
• Applying the framework to sparse-view computed tomography (CT) for fat mass quantification and radiotherapy planning.

Main Results:

• The proposed framework generates bounds with superior semantical interpretation compared to traditional pixel-based methods.
• It enables the identification of plausible-looking but statistically unlikely outlier reconstructions.
• Demonstrated utility in sparse-view CT for clinical applications like fat mass quantification and radiotherapy planning.

Conclusions:

• The framework provides statistically guaranteed, interpretable prediction bounds for deep learning-based image reconstruction.
• It enhances the reliability of clinical assessments derived from sparse-view CT.
• The method effectively flags potentially dangerous outlier reconstructions, improving patient safety.