Conformal Lesion Segmentation for 3D Medical Images

TL;DR

The paper tackles the lack of statistical guarantees in 3D lesion segmentation by introducing Conformal Lesion Segmentation (CLS), which uses split conformal prediction to enforce a test-time FNR constraint. CLS defines an FNR-specific loss to derive per-sample critical thresholds on a calibration set and then uses their approximate $1-\alpha$ quantile as the test-time threshold $\hat{t}$, ensuring $\Pr\left( \mathrm{L}^{\text{FNR}}_{\text{test}}(\hat{t}) \le \varepsilon \right) \ge 1-\alpha$. The method is validated across six 3D-LS datasets and five backbone models, showing substantial FNR reductions compared to fixed thresholds, improved empirical coverage (ECR), and a new prediction-compactness metric to assess spatial precision under risk constraints. CLS remains robust across calibration-test splits and supports practical deployment of uncertainty-aware segmentation in clinical settings by providing a principled, model-agnostic calibration framework with rigorous FNR control.

Abstract

Medical image segmentation serves as a critical component of precision medicine, enabling accurate localization and delineation of pathological regions, such as lesions. However, existing models empirically apply fixed thresholds (e.g., 0.5) to differentiate lesions from the background, offering no statistical guarantees on key metrics such as the false negative rate (FNR). This lack of principled risk control undermines their reliable deployment in high-stakes clinical applications, especially in challenging scenarios like 3D lesion segmentation (3D-LS). To address this issue, we propose a risk-constrained framework, termed Conformal Lesion Segmentation (CLS), that calibrates data-driven thresholds via conformalization to ensure the test-time FNR remains below a target tolerance $\varepsilon$ under desired risk levels. CLS begins by holding out a calibration set to analyze the threshold setting for each sample under the FNR tolerance, drawing on the idea of conformal prediction. We define an FNR-specific loss function and identify the critical threshold at which each calibration data point just satisfies the target tolerance. Given a user-specified risk level $α$, we then determine the approximate $1-α$ quantile of all the critical thresholds in the calibration set as the test-time confidence threshold. By conformalizing such critical thresholds, CLS generalizes the statistical regularities observed in the calibration set to new test data, providing rigorous FNR constraint while yielding more precise and reliable segmentations. We validate the statistical soundness and predictive performance of CLS on six 3D-LS datasets across five backbone models, and conclude with actionable insights for deploying risk-aware segmentation in clinical practice.

Figures (7)

• Figure 1: Illustration of 3D-LS tasks.
• Figure 2: Overview of the CLS framework.
• Figure 3: (a) The output probability matrices contain a substantial number of voxels with predicted probabilities below 0.5. Lesion-region probabilities are relatively uniformly distributed, causing nearly half of ground-truth lesion voxels to be missed under a fixed threshold of 0.5. (b) The FNR-specific loss is monotonically non-increasing with respect to $t$.
• Figure 4: Test-time ECR results on six datasets utilizing five pretrained segmentation models.
• Figure 5: Distribution of FNR-specific loss before (fixed threshold at $0.5$) and after CLS calibration ($\varepsilon = 0.4$, $\alpha = 0.2$), illustrating the reduction in average loss and variance.