Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Conformal Risk Control for Semantic Uncertainty Quantification in Computed Tomography

Jacopo Teneggi, J Webster Stayman, Jeremias Sulam

TL;DR

This work addresses the need for uncertainty quantification in CT reconstruction that is both statistically valid and clinically interpretable. It combines quantile regression with conformal risk control (CRC) and extends it to high-dimensional outputs via K-CRC, aiming to minimize interval length while preserving coverage. The core contribution is sem-CRC, which introduces instance-dependent, organ-specific segmentation to tailor uncertainty to semantic structures, enabling per-organ risk control (and the variant per organ) with guaranteed coverage at level $\epsilon$. Experimental results on CT denoising and reconstruction tasks using TotalSegmentator and FLARE23 demonstrate that sem-CRC produces the tightest, anatomically meaningful uncertainty maps and supports organ-wise risk settings for safer clinical deployment.

Abstract

Uncertainty quantification is necessary for developers, physicians, and regulatory agencies to build trust in machine learning predictors and improve patient care. Beyond measuring uncertainty, it is crucial to express it in clinically meaningful terms that provide actionable insights. This work introduces a conformal risk control (CRC) procedure for organ-dependent uncertainty estimation, ensuring high-probability coverage of the ground-truth image. We first present a high-dimensional CRC procedure that leverages recent ideas of length minimization. We make this procedure semantically adaptive to each patient's anatomy and positioning of organs. Our method, sem-CRC, provides tighter uncertainty intervals with valid coverage on real-world computed tomography (CT) data while communicating uncertainty with clinically relevant features.

Conformal Risk Control for Semantic Uncertainty Quantification in Computed Tomography

TL;DR

This work addresses the need for uncertainty quantification in CT reconstruction that is both statistically valid and clinically interpretable. It combines quantile regression with conformal risk control (CRC) and extends it to high-dimensional outputs via K-CRC, aiming to minimize interval length while preserving coverage. The core contribution is sem-CRC, which introduces instance-dependent, organ-specific segmentation to tailor uncertainty to semantic structures, enabling per-organ risk control (and the variant per organ) with guaranteed coverage at level . Experimental results on CT denoising and reconstruction tasks using TotalSegmentator and FLARE23 demonstrate that sem-CRC produces the tightest, anatomically meaningful uncertainty maps and supports organ-wise risk settings for safer clinical deployment.

Abstract

Uncertainty quantification is necessary for developers, physicians, and regulatory agencies to build trust in machine learning predictors and improve patient care. Beyond measuring uncertainty, it is crucial to express it in clinically meaningful terms that provide actionable insights. This work introduces a conformal risk control (CRC) procedure for organ-dependent uncertainty estimation, ensuring high-probability coverage of the ground-truth image. We first present a high-dimensional CRC procedure that leverages recent ideas of length minimization. We make this procedure semantically adaptive to each patient's anatomy and positioning of organs. Our method, sem-CRC, provides tighter uncertainty intervals with valid coverage on real-world computed tomography (CT) data while communicating uncertainty with clinically relevant features.

Paper Structure

This paper contains 9 sections, 1 theorem, 14 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Quantile regression.
  4. Conformal risk control (CRC).
  5. High-dimensional risk control.
  6. Semantic Uncertainty Quantification
  7. Controlling risk for each organ.
  8. Experiments
  9. Conclusions

Key Result

Proposition 1

For a risk tolerance $\epsilon > 0$, segmentation model $s: {\mathcal{Y}} \to [K]^d$, anchor point ${\widetilde{{\bm \lambda}}}_{{\text{sem}}} \in {\mathbb R}^K_{\geq 0}$, and exchangeable calibration and test points $S_{{\text{cal}}} = \{(X^{(i)},Y^{(i)})\}_{i=1}^{n_{{\text{cal}}}}$, $(X,Y)$, the c

Figures (3)

  • Figure 1: Example calibration data: ground-truth, measurement, and segmented predictions for both tasks and datasets.
  • Figure 2: Example conformalized uncertainty maps on one volume per dataset with each calibration method for the FBP-UNet pipeline. The bottom row shows ${\hat{{\bm \lambda}}}_{{\text{sem}}}$, the semantic uncertainty parameter learned by our method, ${sem\text{-}{\text{CRC}}}$.
  • Figure 3: Mean interval length and risk stratified by organ for the FBP-UNet task across all calibration procedures and datasets. ${\overline{sem}\text{-}{\text{CRC}}}$ is the only procedure that guarantees risk control for each organ.

Theorems & Definitions (2)

  • Proposition 1
  • proof