Study participants

This single-center retrospective study was approved by our Institutional Review Board (IRB ChiECRCT-20170087). A total of 104 orbits from 59 DON patients were included from Tongji Hospital of Huazhong University of Science and Technology (Wuhan, China) between August 2017 and 2020 December. One hundred thirty-one orbits from 69 TAO without DON patients who were treated at Tongji Hospital were also included. They were age-matched with patients with DON. The inclusion and criteria of participants are as follow: (1) age > 18 years, (2) clear refractive media that allowed sufficient image quality, and (3) no treatment with systemic glucocorticoids for at least 3 months before the study. The exclusion criteria of participants are as follow: (1) poor MR image quality; (2) patients with other neurological or ophthalmologic diseases that could explain the vision loss; (3) patients with signs of severe corneal exposure; and (4) patients who had undergone steroid therapy, radiotherapy, or surgical decompression; (5) TAO without DON patients whose age did not match with DON patients.

Thyroid-associated ophthalmopathy (TAO) was diagnosed based on the criteria, established in 1995 by Bartley et al. Dysthyroid optic neuropathy (DON) was made on the basis of the presence of two or more of the following clinical findings: relative afferent pupillary defect when unilaterally affected, color visual defect, reduced visual acuity, papilloedema, visual field defect, and abnormal pattern visual evoked potential test.

Participants were randomly divided into the development and validation sets in an 8:2 ratio. Eighty-three orbits from 44 DON patients and 80 orbits from 40 TAO without DON patients were allocated to the development sets, and the subsequent orbits from participants were allocated to the validation sets. Table 1 shows the detailed demographic information of the participants.

Table 1 Demographic and clinical characteristics of the patients in the training and validation cohortsMRI data acquisition

All participants underwent MRI within one week of the clinical diagnosis of DON or TAO. All MRI scans were conducted on a 3.0 T MR machine (Discovery 750; GE Healthcare) using a 32-channel head coil. The patients were placed in the head-first supine position with their eyes closed during the MRI scanning [11]. Coronal IDEAL-T2WI and axial IDEAL-T2WI sequences were obtained, providing water images. The detailed parameters of coronal and axial IDEAL-T2WI sequences are provided in Table 2.

Table 2 Axial and coronal IDEAL-T2WI parameters

The conventional MRI assessments were independently conducted by two radiologists (two trained neuroradiologists with more than 5 years’ experience in diagnosing DON). They were blinded to the diagnoses (TAO with DON or TAO without DON). The conventional MRI assessment included:

1.

Apical crowding: Crowding of the optic nerve at the orbital apex by enlarged extraocular muscles. It was evaluated on coronal water-IDEAL-T2WI imaging as described by Nugent et al. [19]. (Additional file 1: Fig. S1).

2.

Optic nerve stretching: Globe protrusion exceeds the interzygomatic line at least 21 mm. It was evaluated on axial water-IDEAL-T2WI images as described by Rutkowska-Hinc et al. [9].

3.

Muscle index (MI): Assessing the ratio of the extraocular muscle meridian to orbital meridian. Using the method described by Lynn Barret et al. [20]. (Additional file 1: Fig. S2)

VOI delineation and radiomics feature extraction

The VOI (volume of interest) of the optic nerve tissue was manually segmented with ITK-SNAP software (v. 3.6.0; www.itksnap.org) on the coronal water-IDEAL-T2WI images. VOI was drawn on each image slice that contained the optic nerve tissue by two radiologists with more than 5 years of experience in diagnosing DON independently. The radiologists were unaware of the results of diagnoses (TAO with or without DON). The details of radiomics feature extraction are shown in the Supplementary Material (Additional file 1: Appendix E1) (Fig. 1).

Fig. 1

Manual segmentation of the optic nerve tissue in the participants

Construction of the radiomic signature

Spearman’s correlation coefficient was applied to calculate the relevance and redundancy of the features. The maximum relevance minimum redundancy (mRMR) algorithm was applied to assess the features. The least absolute shrinkage and selection operator (LASSO) logistic regression with tenfold cross-validation was applied to identify the optimal predictive DON features in the development set. The radiomics score (Rad-score) was built using the selected features, weighted by LASSO logistic regression model.

Development of radiomics nomogram and the performance of different models

The conventional MRI prediction model was developed using logistic regression. Then, a multivariate logistic regression analysis included the independent conventional MRI factors. The Rad-score was selected to build the radiomics nomogram model. The calibration curve and Hosmer–Lemeshow test were applied to investigate the performance characteristics of the radiomics nomogram. Receiver operating characteristic (ROC) curves were constructed to compare the performances of each model for differentiating DON from TAO without DON. The specificity, sensitivity, accuracy, and area under the curve (AUC) of each model were calculated. The AUCs were compared by using Delong’s test. Decision curve analysis (DCA) was applied to determine the clinical usefulness of the radiomics nomogram and the conventional MRI evaluation model by quantifying the net benefits for different threshold probabilities in the validation set.

Statistical analysis

Statistical analysis was conducted with R version 3.6.1 and MedCalc version 12.7.0. Interrater and intrarater reliability was assessed by using intraclass correlation coefficient (ICC). p < 0.05 were considered as statistical significance level. The packages of R3.6.1 that were used are shown in the Supplementary Material (Additional file 1: Appendix E2).