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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ophthalmology

Background:

• Thyroid eye disease (TED) assessment relies on accurate segmentation of orbital structures from multi-sequence MRI.
• Current methods face challenges in computational complexity, segmentation accuracy, and integrating diverse MRI sequences.

Purpose of the Study:

• Develop an efficient deep learning framework (RQNet) for precise 3D orbital MRI segmentation.
• Enable robust assessment of TED activity by integrating multi-sequence MRI features.

Main Methods:

• Proposed RQNet, a U-shaped 3D segmentation network with a novel Refined Query Transformer Block (RQT Block).
• Implemented Refined Attention Query Multi-Head Self-Attention (RAQ-MSA) to reduce attention complexity.
• Integrated radiomics features from T1WI, T1CE, and T2WI MRI sequences for TED activity classification using SVM, RF, and LR models.

Main Results:

• RQNet achieved Dice Similarity Coefficients of 83.34-87.15% on TED datasets, outperforming state-of-the-art models.
• The radiomics fusion pipeline yielded AUC values of 84.65-85.89% for TED activity assessment.
• Multi-sequence MRI feature fusion significantly enhanced TED assessment accuracy.

Conclusions:

• RQNet provides an efficient and accurate deep learning solution for 3D orbital MRI segmentation in TED.
• The framework enables robust, radiomics-based assessment of TED activity.
• Integration of multi-sequence MRI features improves diagnostic capabilities for TED.