Interpretable Probabilistic Medical Image Segmentation via Gaussian Process with Explicit Modelling of Annotation Bias and Variability

Researchers propose a novel Gaussian Process-based framework for medical image segmentation that explicitly models annotation bias and variability across multiple raters rather than encoding them implicitly. The approach improves uncertainty calibration in probabilistic predictions while maintaining segmentation accuracy, with quantifiable parameters reflecting individual annotator behavior.

This research addresses a fundamental challenge in medical AI: training robust segmentation models when human annotations inherently contain systematic differences between raters. Traditional deep learning approaches treat annotation variability as noise, while multi-rater probabilistic methods embed annotator characteristics into latent representations that resist interpretation. The proposed logit-space Gaussian Process framework inverts this paradigm by decomposing predictions into an image-dependent reference distribution plus explicit annotator-specific perturbations characterized by bias and variance parameters.

The advancement matters because interpretability directly impacts clinical adoption. When a model's predictions depend on understood, measurable annotator effects, clinicians can better assess prediction reliability and identify systematic annotation patterns. This transparency is essential for regulatory compliance and building trust in AI-assisted diagnostics. The explicit parameterization allows researchers to quantitatively measure which annotators introduce systematic bias versus random variability, enabling targeted improvements in data collection protocols.

For the broader medical AI ecosystem, this work demonstrates how probabilistic frameworks can balance mathematical rigor with practical interpretability. The method's ability to improve uncertainty calibration—a critical metric for clinical deployment—while preserving accuracy suggests scalability to production systems. The public code release amplifies impact by enabling independent validation and adoption across institutions.

Looking ahead, the framework's applicability extends beyond segmentation to classification and detection tasks where multi-annotator datasets are standard. Research should explore whether learned annotator parameters transfer across anatomies or imaging modalities, and how this approach performs with extremely high annotator counts typical in crowdsourced medical labeling.

• →Explicit modeling of annotator bias and variance improves uncertainty calibration in probabilistic medical image segmentation.
• →Gaussian Process framework decomposes predictions into interpretable image-dependent and annotator-specific components.
• →Learned parameters quantitatively reflect individual rater behavior, enabling systematic analysis of annotation patterns.
• →Method maintains competitive segmentation accuracy while enhancing model transparency for clinical adoption.
• →Publicly available code facilitates reproducibility and broader adoption across medical AI research institutions.