CA-WCP: class-aware asymmetric conformal intervals for 3D medical segmentation under shift

Weighted conformal prediction with directional, class-specific bounds keeps nominal 90% coverage on BraTS 2020 while narrowing intervals by 8 to 14%.

Primary source: arXiv:2610.09392: https://arxiv.org/abs/2610.09392

What’s new: Foundation segmentation models such as MedSAM are deterministic and lack calibrated uncertainty under distribution shift. Existing conformal methods are often applied in 2D and assume symmetric errors, which misses the class-specific biases of 3D multi-class segmentation. Quantifying Volumetric Risk: Class-Aware Asymmetric Weighted Conformal Prediction for 3D Medical Image Segmentation (arXiv:2610.09392, 7 Oct 2026) proposes CA-WCP, which combines latent-space density-ratio weighting for covariate shift with directional quantiles for lower and upper volume bounds. Each bound is scaled by a class-specific asymmetry factor derived from validation-set false-positive and false-negative rates.

The authors prove that CA-WCP retains the weighted-exchangeability marginal coverage guarantee for every class. On 3D brain tumor segmentation (BraTS 2020) and a synthetic multi-organ CT benchmark built under covariate shift, the 95% Clopper-Pearson interval for observed coverage contains the nominal 90% level for every semantic class, while interval width falls by 8 to 14% relative to symmetric weighted conformal prediction. The calibrated intervals are also encoded into structured prompts for a multimodal LLM to produce uncertainty-conditioned radiology reports.

Why it matters: Volume estimates drive treatment monitoring, and under-segmentation and over-segmentation have different clinical costs. Directional bounds with a formal coverage guarantee under shift are a practical step toward uncertainty that clinicians can act on, rather than a single point estimate.

Caveats: The guarantee is marginal per class, not conditional on an individual patient or scan. The multi-organ CT evaluation is synthetic, and the real-data evidence comes from one brain tumor dataset. The report-generation component is a demonstration and was not evaluated for clinical accuracy or reader benefit.