Refined query network (RQNet) for precise MRI segmentation and robust TED activity assessment

Abstract Objective: To develop an efficient deep learning framework for precise 3D segmentation of complex orbital structures in multi-sequence MRI and robust assessment of thyroid eye disease (TED) activity, thereby addressing limitations in computational complexity, segmentation accuracy, and integration of multi-sequence features to support clinical decision-making.
Approach: We propose RQNet, a U-shaped 3D segmentation network that incorporates the novel Refined Query Transformer Block (RQT Block) with Refined Attention Query Multi-Head Self-Attention (RAQ-MSA). This design reduces attention complexity from O(N²) to O(N·M) (M\ \ll N) through pooled refined queries. High-quality segmentations then feed into a radiomics pipeline that extracts features per region of interest—including shape, first-order, and texture descriptors. The MRI features from the three sequences—T1-weighted imaging (T1WI), contrast-enhanced T1-weighted imaging (T1CE), and T2-weighted imaging (T2WI)—are subsequently integrated, with support vector machine (SVM), random forest (RF), and logistic regression (LR) models employed for assessment to distinguish between active and inactive TED phases.
Main Results: RQNet achieved Dice Similarity Coefficients of 83.34–87.15% on TED datasets (T1WI, T2WI, T1CE), outperforming state-of-the-art models such as nnFormer, UNETR, SwinUNETR, SegResNet, and nnUNet. The radiomics fusion pipeline yielded area under the curve (AUC) values of 84.65–85.89% for TED activity assessment, surpassing single-sequence baselines and confirming the benefits of multi-sequence MRI feature fusion enhancements.
Significance: The proposed RQNet establishes an efficient segmentation network for 3D orbital MRI, providing accurate depictions of TED structures, robust radiomics-based activity assessment, and enhanced TED assessment through multi-sequence MRI feature integration.