Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Conformal Volume Estimation in 3D Medical Images

Benjamin Lambert, Florence Forbes, Senan Doyle, Michel Dojat

TL;DR

This work tackles uncertainty quantification in 3D medical image volumetry when covariate shifts violate exchangeability in standard conformal prediction. It introduces Weighted Conformal Prediction (WCP) that reweights calibration samples using density ratios, and facilitates efficient density-ratio estimation by leveraging latent representations from a multi-head segmentation network. The method is evaluated on synthetic data with controlled shifts and real BraTS brain tumor MRI data, showing that latent-based WCP recovers target coverage under moderate covariate shifts and maintains calibrated intervals, albeit with wider widths, while Standard CP struggles under shift. The approach provides a scalable path to calibrated volumetric intervals in clinical workflows, with public code availability for replication and adoption.

Abstract

Volumetry is one of the principal downstream applications of 3D medical image segmentation, for example, to detect abnormal tissue growth or for surgery planning. Conformal Prediction is a promising framework for uncertainty quantification, providing calibrated predictive intervals associated with automatic volume measurements. However, this methodology is based on the hypothesis that calibration and test samples are exchangeable, an assumption that is in practice often violated in medical image applications. A weighted formulation of Conformal Prediction can be framed to mitigate this issue, but its empirical investigation in the medical domain is still lacking. A potential reason is that it relies on the estimation of the density ratio between the calibration and test distributions, which is likely to be intractable in scenarios involving high-dimensional data. To circumvent this, we propose an efficient approach for density ratio estimation relying on the compressed latent representations generated by the segmentation model. Our experiments demonstrate the efficiency of our approach to reduce the coverage error in the presence of covariate shifts, in both synthetic and real-world settings. Our implementation is available at https://github.com/benolmbrt/wcp_miccai

Robust Conformal Volume Estimation in 3D Medical Images

TL;DR

This work tackles uncertainty quantification in 3D medical image volumetry when covariate shifts violate exchangeability in standard conformal prediction. It introduces Weighted Conformal Prediction (WCP) that reweights calibration samples using density ratios, and facilitates efficient density-ratio estimation by leveraging latent representations from a multi-head segmentation network. The method is evaluated on synthetic data with controlled shifts and real BraTS brain tumor MRI data, showing that latent-based WCP recovers target coverage under moderate covariate shifts and maintains calibrated intervals, albeit with wider widths, while Standard CP struggles under shift. The approach provides a scalable path to calibrated volumetric intervals in clinical workflows, with public code availability for replication and adoption.

Abstract

Volumetry is one of the principal downstream applications of 3D medical image segmentation, for example, to detect abnormal tissue growth or for surgery planning. Conformal Prediction is a promising framework for uncertainty quantification, providing calibrated predictive intervals associated with automatic volume measurements. However, this methodology is based on the hypothesis that calibration and test samples are exchangeable, an assumption that is in practice often violated in medical image applications. A weighted formulation of Conformal Prediction can be framed to mitigate this issue, but its empirical investigation in the medical domain is still lacking. A potential reason is that it relies on the estimation of the density ratio between the calibration and test distributions, which is likely to be intractable in scenarios involving high-dimensional data. To circumvent this, we propose an efficient approach for density ratio estimation relying on the compressed latent representations generated by the segmentation model. Our experiments demonstrate the efficiency of our approach to reduce the coverage error in the presence of covariate shifts, in both synthetic and real-world settings. Our implementation is available at https://github.com/benolmbrt/wcp_miccai
Paper Structure (11 sections, 6 equations, 3 figures, 2 tables)

This paper contains 11 sections, 6 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Conformal Prediction for volumetry in medical images
  3. Problem definition
  4. Predictive Interval computation using a multi-head segmentation architecture
  5. Weighted Conformal Prediction to tackle covariate shift
  6. Efficient density ratio estimation using latent representations
  7. Experiments
  8. Synthetic dataset with controlled covariate shift
  9. Real-world covariate shift in brain tumor segmentation tasks
  10. Experimental Protocol and Metrics
  11. Results and Discussion

Figures (3)

  • Figure 1: Illustration of the proposed framework. A multi-head segmentation model predicts three distinct masks for each label: a restrictive mask associated with the lower bound volume (red), a permissive mask associated with the upper bound volume (blue), and a balanced mask for the average volume (green). For Weighted Conformal Prediction, a compressed latent representation is extracted from the penultimate convolution filter.
  • Figure 2: Left: Examples of synthetic images with varying Signal-to-Noise ratios (SNRs) and associated ground truths. Right: Distribution of SNRs in the in-distribution and shifted synthetic datasets.
  • Figure 3: Weights of calibration samples estimated by W-Oracle and W-Latent, with and without covariate shift, according to the value of the covariate.