Gaussian Context‐Guided Expert Personalization for Multi‐Rater Medical Image Segmentation

Abstract

Accurate medical image segmentation is often hindered by ambiguity arising from both image quality limitations and variability in expert interpretation. Although datasets with multiple annotators provide richer information about such uncertainty, most existing methods either compress these annotations into a single target, thereby ignoring individual differences, or generate multiple predictions without maintaining alignment with specific experts. As a result, current approaches fail to fully exploit the structure of multi‐rater data. We present a unified framework that simultaneously models segmentation variability and expert‐specific behavior within a single architecture, enabling personalized predictions while preserving diversity. The proposed method learns a shared latent representation that captures the range of plausible annotations and leverages it to generate both diverse and expert‐aligned segmentations. To achieve this, we introduce a Gaussian context‐guided attention mechanism that adaptively extracts relevant features from the latent space in a structured and parameter‐efficient manner. This design allows the model to reflect distinct annotation patterns across experts without relying on heavily parameterized attention modules. By jointly modeling shared anatomical knowledge and expert‐specific variations, the framework maintains consistency across predictions while adapting to individual annotator preferences. We evaluate the proposed approach on the LIDC‐IDRI dataset and a nasopharyngeal carcinoma (NPC‐170) dataset. The results demonstrate consistent improvements over existing methods in both diversity and personalization metrics. Furthermore, the model produces expert‐aligned predictions with enhanced interpretability and reduced computational overhead, suggesting potential applicability in clinical workflows where understanding annotation variability is essential.