CARE: Confidence-aware Ratio Estimation for Medical Biomarkers

Jiameng Li, Teodora Popordanoska, Aleksei Tiulpin, Sebastian Gruber, Frederik Maes, Matthew B. Blaschko

Keywords: medical segmentation, uncertainty quantification, calibration error

TL;DR: This paper provides adaptive and tunable confidence intervals for ratio-based biomarkers, based on pretrained segmentation network.

Abstract:

Ratio-based biomarkers -- such as the proportion of necrotic tissue within a tumor -- are widely used in clinical practice to support diagnosis and treatment planning. In automated clinical workflows, these biomarkers are typically estimated from segmentation outputs by computing region-wise ratios. However, the pointwise estimate captures no uncertainty measurement.
To address this, we propose CARE, a confidence-aware ratio estimation framework considering the error propagation in the segmentation-to-biomarker pipeline. Specifically, we leverage tunable parameters to control the confidence level of the derived bounds. Experiments show that our method produces statistically sound confidence intervals, with tunable confidence levels, enabling more trustworthy application of predictive biomarkers in clinical workflows.

Submission Number: 10