基于MRI的自动分割与多参数定量评估甲状腺眼病眶软组织受累

OBJECTIVES: To develop a deep learning-based multimodal framework for automated segmentation of orbital soft tissues and identify quantitative imaging biomarkers for precise grading of thyroid eye disease (TED).

MATERIALS AND METHODS: This retrospective multicenter study enrolled 330 TED patients from a primary center for model development and 113 patients from two external centers for validation. From the primary cohort, 182 mild and 138 moderate-to-severe TED were subsequently selected for further analysis. All subjects underwent 3 T MRI with water-fat separation and fat-suppressed (FS) T2 mapping sequences. TED-Net, a deep learning model integrating ConvNeXt and Transformer architectures, was developed to segment orbital structures including the extraocular muscles, lacrimal gland, orbital fat, and eyeball. The model automatically extracted both morphological parameters (volume and volume ratio) and functional parameters (water fraction, fat fraction, and FS T2 relaxation time), enabling quantitative comparison between mild and moderate-to-severe TED. Diagnostic performance for evaluating orbital involvement was assessed using receiver operating characteristic analysis and decision curve analysis.

RESULTS: TED-Net achieved Dice similarity coefficients > 0.80 across all orbital structures. Volumetric measurements showed high consistency among different sequences (intraclass correlation coefficient = 0.843). Significant differences in morphological and functional parameters were observed between mild and moderate-to-severe TED (all p < 0.05). The combined volumetric-functional model demonstrated superior diagnostic accuracy (area under the curve [AUC] = 0.982) over the volumetric model (AUC = 0.908).

CONCLUSIONS: TED-Net enables accurate automated segmentation and multiparametric quantification of orbital soft tissues, providing reliable imaging biomarkers for objective assessment of TED severity.

KEY POINTS: Question MRI-based quantitative metrics are essential for assessing orbital involvement in TED, but effective tools for rapid and accurate parameter extraction are currently lacking. Findings TED-Net achieved precise orbital segmentation and multiparametric quantification; the integration of volumetric and functional parameters yielded an AUC of 0.982 for evaluating orbital involvement. Clinical relevance The proposed TED-Net provides a reliable approach for quantifying orbital soft tissue involvement, and its clinical implementation enables more accurate assessment of TED severity.