PatientsThis retrospective study was performed in accordance with the Helsinki Declaration and approved by local Ethics Committee (approval no. UHCT22702). Adult patients (age ≥ 18 years) diagnosed with TAO who underwent orbital 99mTc-DTPA SPECT/CT between January 2021 and August 2022 in two Chinese medical centres were retrospectively analysed. We excluded those with the following characteristics: (1) acute systemic disease and electrolyte disorder; (2) pregnancy or lactation; (3) history of systemic or local hormone therapy, orbital surgery, immunosuppressive therapy, or radiation therapy; (4) other autoimmune, orbital or intracranial disease; (5) other disease that causes exophthalmos; and (6) < 6 months of follow-up. All subjects provided written informed consent. The study flowchart was summarized in Fig. 1. Fig. 1One hundred fifteen patients from Institution 1 were included as the internal cohort and 58 from the Institution 2 were included as the external validation set. Patients in the internal cohort were randomly divided into training and internal testing sets at a ratio of 7:3 (81 patients in the training set and 34 in the internal testing set); this cohort was used for model development and internal testing. The external validation set was used for external validation of the models.Follow-up and treatment evaluationPatient and clinical characteristics, laboratory testing results, treatment, and other data were retrospectively collected and organized through the electronic medical record and picture archiving and communication systems. Recorded data included age, gender, smoking history, history of hyperthyroidism, white blood cell count, neutrophil count, lymphocyte count, alanine aminotransferase, aspartate aminotransferase, free triiodothyronine, free thyroxine, thyroid-stimulating hormone, anti-thyroid peroxidase antibody, thyroid-stimulating hormone receptor antibody, and CAS.Patients with hyperthyroidism and TAO were treated with anti-thyroid drugs to control thyroid function. For patients with moderate-to-severe active TAO, high-dose glucocorticoid pulse therapy was the main treatment; some patients received additional orbital radiotherapy. Certain patients with severe stable disease underwent orbital decompression surgery. The majority of patients with stable TAO didn’t receive any specific treatment and were regularly followed.Patients underwent follow-up examinations once every 3 to 6 months. Routine blood testing and tests for thyroid function and thyroid-stimulating hormone receptor antibody were commonly performed during follow-up, as were determination of CAS and orbital 99mTc-DTPA SPECT/CT or orbital CT when necessary.The treatment efficacy evaluation criteria were as follows: (1) ≥ 2 mm reduction in palpebral fissure; (2) ≥ 2 mm reduction of proptosis; (3) ≥ 2-point decrease in CAS ; (4) ≥ 3 mmHg decrease in intraocular pressure; (5) absence of primary gaze diplopia; and (6) ≥ 10° improvement in eye movement. If three or more criteria were met, the treatment was considered effective; otherwise, it was considered ineffective.
99mTc-DTPA SPECT/CT image acquisition and reconstruction99mTc-DTPA (dose, 370–555 MBq) was injected via an elbow vein in each patient. After 30 to 45 min, SPECT/CT was performed using a Discovery NM/CT 670 Pro system (GE Healthcare, Milwaukee, WI, USA) equipped with a low-energy high-resolution collimator and dual-head detectors. The energy peak was set at 140 keV with a window width of 20% and magnification factor set to 1.00. The imaging matrix was 128 × 128 and the detectors rotated 360° at a rate of 3°/frame. Each frame was collected for 20 s. The CT acquisition parameters were as follows: tube voltage, 140 kV; tube current, 150–300 mA; slice thickness, 1.25 mm; and reconstruction layer thickness, 0.625 mm. Patients were placed in the supine position with their head in the machine’s headrest for the examinations. The bed height was adjusted so the entire head was within the imaging range of the detectors. The orbitoauricular line (the line connecting the external ear aperture and the lateral canthus) was used as the reference baseline. The position of the patient’s head was not moved during imaging.The internal queue was processed using the Volumetrix MI program in the Xeleris 3.0 workstation and then motion correction, fusion correction, attenuation correction, scattering correction, and resolution recovery correction were performed. Next, the ordered subset expectation maximization algorithm was used for reconstruction (2 iterations, 10 subsets), followed by Hann filtering with a cut-off frequency of 0.9. The external queue also used the same algorithm for reconstruction.
99mTc-DTPA SPECT/CT image preprocessingPrior to radiomics feature extraction, all original images were subjected to normalization processes utilizing MATLAB (version 2016a). Given that CT density is relative to technical parameters rather than an absolute value, the images were standardized and normalized. The standardization formula is as follows:$$\:\text{I}\text{m}\text{a}\text{g}\text{e}\:\text{S}\text{t}\text{a}\text{n}\text{d}\text{a}\text{r}\text{d}\text{i}\text{z}\text{a}\text{t}\text{i}\text{o}\text{n}=\frac{x-\mu\:}{Adjusted\:Standard\:Deviation}$$$$\:\text{A}\text{d}\text{j}\text{u}\text{s}\text{t}\text{e}\text{d}\:\text{S}\text{t}\text{a}\text{n}\text{d}\text{a}\text{r}\text{d}\:\text{D}\text{e}\text{v}\text{i}\text{a}\text{t}\text{i}\text{o}\text{n}=\text{m}\text{a}\text{x}({\upsigma\:},\:\frac{1.0}{\sqrt{N}})$$Here, µ represents the mean signal value within the images, x denotes the matrix, σ signifies the standard deviation, and NN is the number of voxels. The normalization equation is$$\:\text{N}\text{o}\text{r}\text{m}\text{a}\text{l}\text{i}\text{z}\text{a}\text{t}\text{i}\text{o}\text{n}=\frac{{\chi\:}_{i}-\text{m}\text{i}\text{n}\left(x\right)}{\text{max}\left(x\right)-\text{m}\text{i}\text{n}\left(x\right)}$$Where xi indicates the signal of a voxel. In alignment with the standardized workflow of radiomics, image preprocessing is a prerequisite step before radiomics feature extraction can occur.
99mTc-DTPA SPECT/CT image quantizationImage quantization refers to the transformation of image grayscale values into a discrete set of grayscale values. Before radiomics feature extraction, the image was quantized with a fixed bin width of 5, a choice based on the example settings of Pyradiomics (https://github.com/AIMHarvard/pyradiomics/tree/master/examples/exampleSettings). According to the protocol, the ideal range for the number of bins is between 16 and 128 bins. A suitable bin width can be determined by extracting a feature known as the first-order range, ensuring it falls within this bin range. The outcomes of this range are detailed in Supplementary file. An absolute discretization was carried out with a fixed bin size (binsize = 5), assigning new bins to pixel intensities for each BS grayscale level starting at 0. The formula is as follows:$$\:\text{I}\text{B}\text{S}\left(x\right)=\left(\frac{I\left(x\right)}{BS}\right)-min\left(\frac{I\left(x\right)}{BS}\right)+1$$Here, I(x)I(x) signifies the intensity of a voxel; BS (binsize) indicates the bin size, and IBS(x)IBS(x) represents the discretized grayscale level of voxel xx (29). All images were then converted into Gaussian and wavelet forms. Wavelet filtering results in eight decompositions at each level. The wavelet filter settings are as follows: (1) start_level [0]: an integer representing the base level of wavelet from which the first set of decompositions for signature calculation is derived. (2) level [1]: an integer for the number of wavelet decomposition levels used for signature calculation. (3) wavelet [“coif1”]: a string indicating the type of wavelet decomposition. It is an enumerated value validated against possible values present in pyWavelet.wavelist. The current possible values are for pywavelet version 0.4.0. The Gaussian image is obtained by convolving the image with the Laplacian of a Gaussian kernel. The Gaussian kernel is calculated using the following formula:$$\text{G}\left({x},y,z,\upsigma\right)=\frac{1}{{(\sigma\:\sqrt{2\pi\:)}}^{3}}{e}^{-\frac{{\mathcal{X}}^{2}+{\mathcal{Y}}^{2}+{\mathcal{Z}}^{2}}{2{\sigma\:}^{2}}}$$Region-of-interest segmentation, image pre-processing and feature extractionsThe maximum cross-sectional area of the superior, inferior, medial, and lateral rectus on both eyes was selected as the region of interest(ROIs), which were segmented on the CT images through the open source 3D-Slicer software version 5.0.2 (https://download.slicer.org/) (Fig. 2). The ROIs of both the internal and external queues were delineated manually by two nuclear medicine physicians with 2–3 years’ experience who were blinded to subject information. All ROIs were confirmed by another nuclear medicine physician who had over 8 years of experience. Fig. 2Radiomics workflow. The term ‘LASSO’ stands for Least Absolute Shrinkage and Selection Operator, a widely used method for regression analysis when dealing with high-dimensional data. LASSO regression selects significant and distinguishable features to construct the radiomics signature, effectively simplifying the model and reducing the risk of overfitting.To accurately map the delineated ROIs to SPECT, MATLAB software version 2016a (MathWorks, Natick, MA, USA) was used to achieve complete spatial matching of slice thickness, number of slices, and inter-slice spacing between the CT and SPECT images in order to ensure accuracy. Before extracting the texture features of the ROIs, all images were standardized, normalized, and discretized according to the Image Biomarker Standardization Initiative. Subsequently, the CT and SPECT images were subjected to Gaussian and wavelet transformations. Then, the voxels were reconstructed to a size of 1 × 1 × 1 mm3. Finally, three sets of images were obtained: original images, Gaussian-transformed images, and wavelet-transformed images.According to the feature guidelines of the Image Biomarker Standardization Initiative, the pyradiomics feature package (github.com/Radiomics/pyradiomics) was used for image feature extraction in Anaconda Prompt software version 4.2.0 (https://www.anaconda.com/download). Six types of texture features were extracted based on the original, Gaussian-transformed, and wavelet-transformed images: first-order, shape, gray-level co-occurrence matrix, gray-level run-length matrix, gray-level size-zone matrix and gray-level distance-zone matrix.Feature selection and radiomics model establishmentTexture features of the training set were used to construct radiomics labels. First, the minimal redundancy maximal relevance algorithm (mRMR) was used to remove redundant and irrelevant features. Then, the least absolute shrinkage and selection operator algorithm was used to build radiomics signature based on the minimum binomial deviation that could distinguish TAO activity were selected (Fig. 2). In the 99mTc-DTPA SPECT/CT images, the Logλ(0.0342) identified 20 features (Fig. 2). These relevant radiomics signatures were wavelet_LLL_firstorder_ 90Percentile. despect, wavelet_ LLL_ glrlm_GrayLevelNonUniformity.ct, wavelet_LLL_glrlm_ GrayLevel NonUniformity.ct, wavelet_LLL_glcm_Imc2.ct, log_sigma_3_0_mm_3D_ glcm_Idmn.ct, wavelet_LLL_firstorder_Maximum.despect, wavelet_LHL_glrlm_ LongRunEmphasis.despect, wavelet_LLL_firstorder_Skewness.ct, wavelet_LLL_ glszm_ZoneVariance.ct, wavelet_HHH_glszm_GrayLevelNonUniformity Normalized. despect, wavelet_LLL_glcm_InverseVariance.despect, original_glcm_ Imc2.ct, wavelet_HHH_glszm_LowGrayLevelZoneEmphasis.ct, log_sigma_3_0_ mm_ 3D_ glszm_SmallAreaLowGrayLevelEmphasis.despect, log_sigma_2_0_mm_ 3D_ glszm_SmallAreaLowGrayLevelEmphasis.ct, wavelet_LLL_glcm_Idmn.despect, original_glrlm_ShortRunLowGrayLevelEmphasis.despect, original_glcm_ Cluster Prominence.ct, wavelet_HHL_glszm_SmallAreaLowGrayLevelEmphasis.ct,Wavelet_HLL_glcm_Imc1.ct (Supplementary file). Rad-score was established by combining the corresponding weights of texture, and the training set patient data were subjected to 10-fold cross-validation. Since this study included both CT and SPECT images, three Rad-scores were constructed, including Rad-scoreCT, Rad-scoreSPECT, and Rad-scoreSPECT/CT .Establishment of clinical models and combined clinical-radiomics modelsFirst, clinical characteristics were selected to built clinical model to analysis, active and inactive disease in the training and internal testing sets. Then, the multivariate logistic regression model was used to built clinical model based the minimum. Akaike information criterion (AIC) were used to analyse the clinical features of the training set. Patient age and CAS were ultimately included in the clinical model construction. In addition, a combined model based Rad-score and retained clinical parameters were built using a multivariate logistic regression model.Model effectiveness evaluationAll constructed models were first validated in the training and internal validation sets. The performance of all radiomics models for determining TAO activity was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC). By stepwise comparison of the statistical differences between the ROC curves of each model, the optimal radiomics model was ultimately selected. The Hosmer–Lemeshow test and calibration curve were used to validate the fitting degree between the predicted values and the true values of the model. Decision curve analysis was used to evaluate the clinical benefits among three models. Finally, the clinical utility of the model for the assessment of TAO activity was validated using the AUC in the external validation set.Statistical analysisStatistical analyses were performed using R software 3.6.4 (www.Rproject.org) and the following statistical packages: tidyverse, caret, pROC, DMWR, rmda, ggpubr, mRMRe, DescTOOLs and irr. Medcalc software (www.medcalc.org) was used to analyse the ROC curves. The differences between ROC curves of the different models were compared using the Delong test. Line graphs were plotted using Prism software (GraphPad, San Diego, CA, USA).
Diagnostic precision in thyroid-associated ophthalmopathy using multi-center radiomics with 99mTc-DTPA SPECT/CT
