Whole-volume ADC Histogram and Texture Analyses of Parotid Glands as an Image Biomarker in Evaluating Disease Activity of Primary Sjögren’s Syndrome

Diffusion weighted imaging (DWI) has proven to be sensitive for detecting early injury to the parotid gland in pSS (primary Sjögren’s syndrome). Here, we explored the application of ADC histogram and texture analyses for evaluating the disease activity of pSS. A total of 55 patients with pSS who met the classification criteria of the 2002 AECG criteria prospectively underwent 3.0-T magnetic resonance imaging (MRI) including DWI (b = 0 and 1000 s/mm2). According to the ESSDAI score, 35 patients were categorized into the low-activity group (ESSDAI < 5) and 20 into the moderate-high-activity group (ESSDAI ≥ 5). Via analysis of the whole-volume ADC histogram, the ADCmean, skewness, kurtosis, and entropy values of the bilateral parotid glands were determined. Multivariate analysis was used to identify independent risk factors for predicting disease activity. The diagnostic performance of the indexes was evaluated via receiver operating characteristic (ROC) analysis. ROC analysis showed that the anti-SSB, lip biopsy, MRI morphology, ADC, ADCmean, and entropy values were able to categorize the disease into two groups, particularly the entropy values. The multivariate model, which included anti-SSB, MRI morphology and entropy, had an area under the ROC curve of 0.923 (P < 0.001). The parotid entropy value distinguished disease activity in patients with pSS, especially combined with anti-SSB and MRI morphology.

Diagnostic performance of various indexes. Univariate analysis showed significant differences for the anti-SSB, lip biopsy, MRI morphology, ADC, ADC mean , and entropy values between the moderate-high-activity and low-activity group (all P < 0.05). ROC analysis showed that the anti-SSB, lip biopsy, MRI morphology, ADC, ADC mean , and entropy values performed well in differentiating the moderate-high-activity group from the low-activity group. Entropy values had the highest area under the ROC curve (AUC) of 0.853 ( Table 2). The multivariate model calibration was assessed with the goodness-of-fit Hosmer-Lemeshow test (P = 0.906). The optimal combination included anti-SSB, MRI morphology and entropy and yielded a sensitivity of 95.0%, specificity of 77.1%, accuracy of 83.6%, and AUC of 0.927 (P < 0.001) (  Table 1. The parotid magnetic resonance imaging (MRI) indexes of patients with primary Sjögren's syndrome having different activities. Note, ADC, the mean apparent diffusion coefficient (ADC) value, obtained from one region of interest (ROI) in single slice. ADC mean , the mean value of all ADC values within the volume of interest (VOI). Skewness, the histogram asymmetry degree around the mean. Kurtosis, a measurement of the histogram sharpness. Entropy, the distribution of gray levels within the VOI. *P < 0.05.
Correlation between the ADC histogram and texture parameters and the scores for each of the ESSDAI items. As shown in Appendix Table 1, skewness was negatively correlated with cutaneous injury (r = 0.333, P = 0.013) and positively correlated with serum biomarkers (r = 0.307, P = 0.023). Kurtosis was negatively correlated with constitutional symptoms (r = 0.304, P = 0.024) and positively correlated with serum biomarkers (r = 0.340, P = 0.011). Entropy was negatively correlated with muscular injury (r = 0.268, P = 0.018) and positively correlated with constitutional symptoms (r = 0.318, P = 0.048).

Intra-and interobserver agreements of MRI interpretation.
The intra-and interobserver agreements regarding the evaluation of the MRI morphology grade were excellent (kappa coefficient = 0.895 and 0.950, respectively). The measurements of the parotid ADC value and all of the ADC histogram and texture parameters showed excellent intra-and interobserver agreement, with ICCs ranging from 0.899 to 0.982 (Table 4).

Discussion
This study compared the differences among clinical, laboratory, and imaging indexes between patients with pSS with different levels of disease activity and confirmed the feasibility of using parotid ADC histogram and texture analyses to predict the disease activity of pSS. The ESSDAI score was 3.4 ± 1.9 in the present study cohort, which was slightly higher than 3.14 ± 3.47 in the report by Fidelix et al. 8 . This difference might be because patients with pSS with low disease activity accounted for 63.3% of all patients in the present study, but 72.9% of all patients based on ultrasonography in the study by Fidelix et al. 8 .
This study found that the moderate-high-activity group had positive anti-SSB more frequently compared with the low activity-group. Maslinska et al. 17 reported that the presence of anti-SSB was significantly affected by the higher activity of the disease, which was consistent with the findings of the present study. Additionally, the moderate-high-activity group had positive MRI morphology, X-ray sialography, and lip biopsy findings more , corresponding to the apparent diffusion coefficient (ADC) map, and histogram of the bilateral parotid glands of an 18-year-old woman with low-activity primary Sjögren's syndrome (pSS). Her European League Against Rheumatism (EULAR) pSS disease activity index (ESSDAI) score was 4 (haematological score 2+ serum biomarker score 2). MRI morphology, ANA, anti-SSA, X-ray sialography, and lip biopsy showed positive findings, whereas her anti-SSB and ocular tests were negative. The parotid ADC, ADC mean , skewness, kurtosis, and entropy values were 959.1 × 10 −6 mm 2 /s, 1006.2 × 10 −6 mm 2 /s, −0.177, 4.746, and 5.690, respectively. (D-F) Axial DWI scan (b = 1000 s/mm 2 ), corresponding to the ADC map, and histogram of the bilateral parotid glands in a 66-yearold woman with moderate-activity pSS. Her ESSDAI score was 6 (haematological test, serum biological marker score 2 and constitutional symptoms, articular score 1). MRI morphology, ANA, anti-SSA, anti-SSB, and X-ray sialography showed positive findings, whereas the ocular tests and lip biopsy were negative. The ADC, ADC mean , skewness, kurtosis, and entropy values were 943.0 × 10 −6 mm 2 /s, 1015.6 × 10 −6 mm 2 /s, 0.359, 5.254, and 6.181, respectively. Note the dashed lines covering the edge of the right parotid glands.
ScieNTific REPoRTS | (2018) 8:15387 | DOI:10.1038/s41598-018-33797-x frequently compared with the low-activity group, which has not been previously reported. We speculate that the parotid and lip glands are more prone to injury in patients with pSS with high disease activity.
The parotid ADC values from one ROI and ADC mean based on the whole-volume histogram and texture analyses were significantly lower in the moderate-high-activity group than in the low-activity group. Previous studies have reported that the parotid ADC value increased during the early stages of pSS due to oedema and the increased vascular permeability of parotid glands 3,18 , but it decreased during the late stages of the disease due to fatty deposition and atrophy of the parotid glands 19,20 . The decreased parotid ADC value in the moderate-high-activity group may also be involved in the decreased microvascular perfusion of the glands. The present study failed to detect a significant difference in the parotid volume between different levels of disease activities.
We found that skewness and kurtosis correlated with some of the ESSDAI items. A lower skewness indicates a higher frequency of high ADC values, which may be due to inflammation and micro-necrosis. A lower kurtosis indicates high heterogeneity of the tissue, which may be due to cell proliferation and necrosis. The entropy value of the parotid glands was significantly higher in the moderate-high-activity group than in the low-activity group. It was speculated that the entropy value might be related to the inflammation characteristics of parotid injury. Texture analysis reflects inflammation disease activity based on the histological characteristics, including transmural inflammation, fissuring ulcers, and oedema 15,21 . Makanyanga et al. 15 found that entropy reflected Crohn's disease activity according to the microenvironment heterogeneity and complexity. Chu et al. 22 reported that the parotid gland microenvironmental complexity increased with aggravation of the injury grade.

Index
Cutoff value Sensitivity Specificity Accuracy AUC P   Parotid entropy could distinguish pSS with moderate-high disease activity from pSS with low disease activity with an AUC of 0.853, which was higher than any other single index, including clinical, laboratory, and other imaging parameters. The multivariate model that included parotid entropy, anti-SSB, and MRI morphology yielded a higher value on the goodness-of-fit Hosmer-Lemeshow test. The AUC of this model reached 0.927, with a high sensitivity of 95.0% and relatively low specificity of 77.1%. The disease activity of pSS could be reflected more comprehensively and accurately by combining the information on the parotid glands, serum biomarkers, and lip biopsy.
The fat content and heterogeneity of the parotid glands increase with age, which might affect the DWI-derived parameters. In our study, there was no significant difference in age between the moderate-high-activity and low-activity groups.
Nevertheless, via imaging, the injury has been observed to be significantly advanced in patients with low disease activity and a well-established long-standing disease. Therefore, it is critical to determine the duration of the disease because it is an important confounding factor. Hence, we reviewed the disease duration of all pSS patients. Nevertheless, the disease duration showed no significant differences between the low and moderate-high activity groups. Fidelix et al. 8 also reported no significant correlation between disease duration and ESSDAI, which is consistent with our findings.
Limitations. The present study had several limitations. First, the sample size was relatively small, although it was larger than previous MRI studies on pSS 10,12,23 . Second, other functional MRI analyses, such as dynamic contrast-enhanced imaging, were not performed or compared with DWI. Third, the diagnostic criteria proposed in our study should be validated in another cohort. Fourth, parotid biopsy was not performed due to its invasiveness and patient discomfort. These issues require further investigation.

Conclusion
In conclusion, the entropy value derived from whole-volume ADC histogram and texture analyses of the parotid glands shows great potential for predicting the disease activity of pSS. The diagnostic performance of whole-volume ADC histogram and texture analyses can be further improved by combining parotid entropy with anti-SSB and the MRI morphology, which can serve as an imaging biomarker of pSS disease activity.

Materials and Methods
Patients. This study was approved by the ethics committee of Nanjing Drum Tower Hospital. Written The exclusion criteria were as follows: (1) a history of radiotherapy applied to the head and neck, hepatitis C virus infection, acquired immunodeficiency disease, lymphoma, sarcoidosis, or drug use, such as diuretics, tricyclic antidepressants, and/or anticholinergic agents causing xerostomia (n = 1); (2) a diagnosis of secondary SS associated with other autoimmune diseases, such as systemic lupus erythematosus and rheumatoid arthritis (n = 5); or (3) contraindications to MRI, such as a cardiac pacemaker or artificial cochlear implantation (n = 1).
The two groups significantly differed at the beginning of the study in the glandular domain of the ESSDAI, which was positive in only one patient (2.8%, 1/36) in the low-ESSDAI group and in 31.0% (9/29) rather than 47.7% of patients in the moderate-high ESSDAI group, which was close to the positive rate of 28.12% in Seror R et al. 's report 2 in pSS patients. A Chi-square test showed significant differences in the glandular domain of ESSDAI between the two groups (P = 0.002). ROC analysis showed that the glandular involvement of ESSDAI was able to differentiate the moderate-high-activity group from the low-activity group with a low sensitivity of 31.0%, high specificity of 97.2% and accuracy of 67.7%. To rule out the interference of the glandular domain, we excluded pSS patients with glandular involvement of ESSDAI. Finally, 55 patients were enrolled (53 females and 2 males; mean age, 46.8 ± 14. pSS disease activity index evaluation. Two rheumatologists (C.W. and H.Y.Z., with 8 and 10 years of experience, respectively) evaluated the pSS disease activity according to the ESSDAI score criteria, including information on constitutional symptoms; lymphadenopathy; glandular, articular, cutaneous, pulmonary, renal, muscular, and peripheral nervous and central nervous system injury; and haematological and serum biomarkers, based on clinical, laboratory, and radiological examinations. Each system was scored as 0 for no activity, 1 for low activity, 2 for middle activity, and 3 for high activity 2 . A third rheumatologist (L.Y.S., with 20 years of experience) was consulted in the case of any inconsistency between the two primary rheumatologists. The ESSDAI score was the sum of 12 systems, which was defined as low (ESSDAI < 5), moderate (5 ≤ ESSDAI ≤ 13), or high (ESSDAI ≥ 14) activity levels 6 . In the present study, patients with low activity served as group 1 and those with moderate or high activity served as group 2. The mean ESSDAI score of the 55 patients was 3.4 ± 1.9 (0-7), in-plane resolution = 1.8 × 2.0 mm 2 ; acquisition matrix = 132 × 120; reconstruction matrix = 256 × 256; NSA = 4; bandwidth = 819.9 Hz/pixel). Volume shim and high order shim were applied to counter an inhomogeneous B0 field. A DWI with a short TI inversion recovery (STIR) fat suppression was applied and combined with slice-selective gradient reversal (SSGR) to further improve fat suppression 25 . The b value used in our article was as reported previously (b value = 0, 1000 s/mm 2 ) 10,12 . Three motion-probing gradients along the readout, phase-encoding, and slice-selection directions were used. The diffusion registration adopted a 3D affine registration using the "local correlation (LC)" algorithm. The LC algorithm gave a similar or better performance compared with the mutual information algorithm 26 , which avoided various types of subjective movements as well as eddy current effects. The acquisition time of DWI was approximately 3 min 48 s, and the total scan time was approximately 17 min 47 s. All participants underwent MRI successfully without any side effects or discomfort.  Medical Systems, Best, The Netherlands). Two radiologists (C.C. and J.H., with 2 and 10 years of experience in head and neck radiology, respectively) who were blinded to the clinical and laboratory information performed the interpretation and measurement independently. The injury degree of the unilateral parotid gland was evaluated based on T1W, T2W, and T2-STIR images according to the scale proposed by Makula et al. 23 as follows: grade 0, normal homogeneous gland parenchyma; grade 1, fine reticular or small nodular structure with the diameter of nodules <2 mm; grade 2, medium nodular pattern with the diameter of nodules 2-5 mm; and grade 3, coarsely nodular with the diameter of nodules >5 mm. Grade 0 was considered to be negative and grades 1-3 to be positive on MRI for diagnosing SS. A consensus was achieved by discussion if any divergence existed between the two radiologists. The ADC maps were generated from DWI scans with the software integrated within the workstation using a monoexponential model: S = S0 × exp(-b × ADC). The DWI scan (b = 1000 s/mm 2 ) showing that the largest slice of the parotid glands was selected, and an ROI was manually drawn to cover the unilateral parotid gland as large as possible (mean, 464.1 ± 107.5 mm 2 ; range, 226.2-723.5 mm 2 ), keeping a distance of 1 mm from the boundary and carefully avoiding the retromandibular vein and external carotid artery within the gland. The ROIs were copied to the ADC maps automatically, and the mean ADC value of the ROI was obtained. The mean value of the bilateral parotid glands was calculated. Whole-volume ADC histogram and texture analyses were performed using in-house software (Image Analyzer 2.0, China), the details of which were described in previous studies 13,[27][28][29] . A series of ROIs were manually drawn to cover the parotid gland as large as possible on each slice of the DWI scan (b = 1000 s/mm 2 ). The ROIs were copied to the ADC maps automatically. After selecting all of the ROIs of the unilateral parotid gland (slice number, 5-10; mean, 8 ± 1), the volume of interest (VOI) was composed (volume range, 1060.2-8927.8 mm 3 ; mean, 3710.6 ± 1442.2 mm 3 ) to obtain the following parameters, which were calculated using the following formulas (1)(2)(3)(4), where X indicates the set of all ADC values, N is the number of sampled ADC pixels, X is the mean of X, and P(i) is the frequency of voxels with intensity i divided by N.
(1) ADC mean , the mean value of all ADC values within the VOI, (2) skewness, the histogram asymmetry degree around the mean, (3) kurtosis, a measurement of the histogram sharpness, (4) entropy, the distribution of grey levels over the VOI, −∑ = The measurements of each radiologist were recorded separately for interobserver agreement analysis. The mean value of the two radiologists was calculated as the final value of each patient. One radiologist (X.X) repeated all of the measurements 1 month later for intraobserver agreement analysis.

Statistical analyses.
Quantitative data showing a normal distribution (Kolmogorov-Smirnov test, all P > 0.05) are presented as the mean ± standard deviation, and qualitative data are presented as ratios. Continuous variables were compared with two independent-samples t tests, and categorical variables were compared with the Chi-square test between the two groups. The Mann-Whitney U test was used to compare the differences in age and disease duration distribution. The diagnostic performance of each index was tested via receiver operating characteristic (ROC) analysis. Cutoff values were established by calculating the maximal Youden index (Youden index = sensitivity + specificity − 1). Moderate-high activity was defined as a positive result. Univariate analysis was used to compare the indexes for differentiating moderate-high-activity from the low-activity group. Binary logistic regression analysis with a backward stepwise selection procedure was performed to identify the independent predictors for differentiating the moderate-high from the low-activity group. Multivariate model calibration was assessed with the goodness-of-fit Hosmer-Lemeshow test and graphical decile group probability through a calibration plot. The McNeil test was used to compare the area under ROC curves (AUCs). The Spearman rank test was used to correlate the ADC histogram parameters and the scores for each of the ESSDAI items. The kappa coefficient was calculated to evaluate the intra-and interobserver agreements in MRI morphology assessment. Intra-and interobserver agreements in the measurement of ADC parameters were estimated by calculating the intraclass correlation coefficients (ICCs) (0.000-0.200, poor; 0.201-0.400, fair; 0.301-0.600, moderate; 0.601-0.800, good; 0.801-1.000, excellent). Statistical analyses were performed using SPSS (version 22.0 for Microsoft Windows x64, SPSS, IL, USA). A two-tailed P value less than 0.05 was considered statistically significant, and a P value less than 0.05 was considered significant in univariate analysis.

Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.