CONReg: Uncertainty-Aware Medical Image Registration Using Conformal Prediction.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Radiology

Background:

• Deep learning (DL) models for medical image registration lack uncertainty quantification, limiting clinical trust.
• Uncertainty quantification (UQ) is crucial for identifying unreliable predictions in medical imaging.

Purpose of the Study:

• To introduce CONReg, a novel framework for voxelwise and case-level UQ in medical image registration.
• To improve the reliability and interpretability of DL-based registration through statistically principled uncertainty estimation.

Main Methods:

• Integrated quantile regression with conformal prediction (CP) to generate predictive intervals for dense displacement fields (DDFs).
• Utilized a 3D U-Net for predicting DDFs and their quantile bounds on brain and lung datasets.
• Developed uncertainty bounding boxes (UBBxs) at keypoints and stratified cases into certain/uncertain groups.

Main Results:

• Achieved empirical coverage between 0.92-0.98, demonstrating reliable UQ.
• Certain keypoints and cases showed significantly lower target registration error (p < 0.05).
• Certain cases exhibited significantly lower mean squared error (MSE) compared to uncertain cases (p < 0.05).

Conclusions:

• CONReg provides a statistically sound method for assessing registration uncertainty.
• The framework enhances the trustworthiness and interpretability of DL models in medical image registration.
• CONReg facilitates better clinical decision-making by highlighting areas of potential prediction unreliability.