Collaborative Feature and Persona Enhancement for Trustworthy Medical Foundation Models

Fanqi Kong, Ziming Zhao

Keywords: persona-aware foundation models; medical image segmentation; fairness and bias; trustworthy medical AI.

TL;DR: We introduce a persona-aware CEIGM-UNet that jointly models medical image segmentation and patient personas, achieving state-of-the-art accuracy with fewer parameters while improving fairness across demographic groups.

Abstract: Visual state-space models (SSMs) such as Mamba have emerged as strong backbones for medical image segmentation due to their ability to capture long-range dependencies with linear computational complexity. However, existing Mamba-based architectures provide limited support for explicit feature selection, targeted feature enhancement, and dedicated multi-scale representation learning, leaving them vulnerable to confusing anatomical structures and imaging noise. Moreover, segmentation models deployed in real-world clinical environments must remain robust across heterogeneous demographic profiles without amplifying spurious or stereotype-like correlations. We introduce Persona-guided Collaborative Feature Enhancement and Inception GroupMamba UNet (Persona-Guided CEIGM-UNet), a Mamba-based segmentation framework that addresses these limitations. Built upon a GroupMamba encoder, our design incorporates: (i) a Collaborative Feature Enhancement Layer (CFEL) that integrates attention-guided refinement, dynamic up-convolution, and multi-scale enhancement gating; (ii) a Modulated Inception Group Mamba Layer (MIGML) that couples multiscale local pattern extraction with long-range dependency modeling; and (iii) a lightweight Demographic-Aware Persona Modulation (DAPM) branch that maps demographic meta-information to bounded channel-wise modulation factors, enabling mild, controlled feature adaptation. Experiments on the Synapse and ACDC datasets show that the CEIGM-UNet backbone achieves state-of-the-art performance with fewer parameters and competitive FLOPs. A preliminary fairness evaluation on Synapse, assessing equalized odds differences and generalized Dice disparities across age and sex, suggests that persona-guided modulation can reduce group-wise performance gaps relative to strong Transformer baselines.

Submission Number: 139