SACP: Spatially-Aware Conformal Prediction in Uncertainty Quantification of Medical Image Segmentation
Keywords: Uncertainty Quantification, Conformal Prediction, Medical Imaging, Image Segmentation, Deep Learning
TL;DR: A novel uncertainty quantification method for medical image segmentation sensitive to critical anatomical interfaces allowing optimal surgery planning.
Abstract: Conformal Prediction provides statistical coverage guarantees for uncertainty quantification but fails to account for spatially varying importance of predictive uncertainty in medical image segmentation. This paper introduces a spatially-aware conformal prediction framework that enhances uncertainty quantification by incorporating spatial context near critical anatomical interfaces such as a vessel or critical organ. Our framework consists of three key components: (1) a base nonconformity score derived from segmentation model probabilities, (2) a calibration mechanism that applies structure-specific importance weights based on spatial proximity, and (3) a prediction set construction method that preserves mathematical coverage guarantees while providing targeted uncertainty quantification in critical regions. The calibration mechanism employs a distance-weighted scoring function that exponentially decays with distance from key interfaces, allowing for structure-specific importance factors and adaptive uncertainty estimation. We develop pooled and domain-specific calibration strategies to handle multi-center variability, enabling robust performance across different imaging protocols and populations. We validate our approach on tumor segmentation in pancreatic adenocarcinoma imaging from two medical centers. Results demonstrate that our method achieves the desired coverage levels while generating prediction sets that adaptively expand near critical interfaces.
Primary Subject Area: Uncertainty Estimation
Secondary Subject Area: Segmentation
Paper Type: Methodological Development
Registration Requirement: Yes
Reproducibility: https://github.com/tailabTMU/SACP
Visa & Travel: Yes
Submission Number: 238
Loading