Step-Attention Refinement of DINOv3 Features for Efficient Anterior Eye Segmentation

TL;DR

The paper introduces a lightweight method that enhances DINOv3's vision transformer features for anterior eye segment segmentation by applying a step-attention refinement module across multi-level features before decoding. This approach improves segmentation accuracy and robustness under domain shifts while using minimal additional parameters. It achieves state-of-the-art performance on a diverse clinical dataset and generalizes well to unseen public datasets.

Why It Matters

Robust and efficient segmentation of the anterior eye is critical for clinical diagnostics and biometric systems, especially when dealing with varied imaging conditions and limited labeled data. This work demonstrates how carefully designed feature refinement can unlock the full potential of foundation models in medical imaging without heavy computational costs.

Key Findings

• The step-attention refinement module effectively adapts multi-scale transformer features for dense prediction tasks with low parameter overhead.
• The method achieves 85.55% mIoU on a private clinical dataset with diverse acquisition protocols and outperforms both convolutional and transformer-based baselines.
• It shows superior generalization across four unseen public datasets, indicating strong resilience to domain shift in real-world clinical deployment.