Efficient Conformal Volumetry for Template-Based Segmentation

TL;DR

ConVOLT is introduced, a CP framework that achieves efficient volumetric UQ by conditioning calibration on properties of the estimated deformation field from template-based segmentation, and calibrates a learned volumetric scaling factor from deformation space features.

Abstract

Template-based segmentation, a widely used paradigm in medical imaging, propagates anatomical labels via deformable registration from a labeled atlas to a target image, and is often used to compute volumetric biomarkers for downstream decision-making. While conformal prediction (CP) provides finite-sample valid intervals for scalar metrics, existing segmentation-based uncertainty quantification (UQ) approaches either rely on learned model features, often unavailable in classic template-based pipelines, or treat the registration process as a black box, resulting in overly conservative intervals when applied directly in output space. We introduce ConVOLT, a CP framework that achieves efficient volumetric UQ by conditioning calibration on properties of the estimated deformation field from template-based segmentation. ConVOLT calibrates a learned volumetric scaling factor from deformation space features. We evaluate ConVOLT on template-based segmentation tasks involving global, regional, and label volumetry across multiple datasets and registration methods. ConVOLT achieves target coverage while producing substantially tighter intervals than output-space conformal baselines. Our work paves way to exploit the registration process for efficient UQ in medical imaging pipelines.

Figures (3)

• Figure 1: Overview of ConVOLT. We show a standard template-based segmentation pipeline for case $i$ (black) where a moving image is registered to a fixed image, resulting in a deformation field. The fixed labels are warped to compute predicted labels and volumes $\widehat{Y}^0_{i,l}$ on the moved image. ConVOLT (blue) trains a label-specific regressor on the training dataset $D_{tr}$ with statistical features extracted from the deformation field features to predict a multiplicative scaling factor $\widehat{\beta}_{i,l}$. The scaling factor is conformalized using the calibration dataset $D_{cal}$ to attain valid prediction intervals.
• Figure 2: ConVOLT achieves high efficiency compared to CQR for label volume guarantees. For $\alpha=0.1$, we show interval size inflation of CQR (best performing baseline) in % for each label in OASIS with ConVOLT as the baseline. We find that ConVOLT achieves higher efficiency on the majority of labels.
• Figure 3: Coefficient magnitudes explain ConVOLT's efficiency improvements. We show Mean$\pm$STD of absolute ridge coefficients for features across 100 random calibration-test splits. Large and stable magnitudes indicate features are consistently used to predict multiplicative correction, while small and unstable magnitudes indicate the limited predictive value of features and explains weaker performance.