Quantitative evaluation of diffusion-weighted MRI for differentiating benign and malignant thyroid nodules larger than 4 cm

The study followed the Declaration of Helsinki (revised 2013). The Institutional Ethics Committee of Minhang Hospital affiliated with Fudan University approved this observational, retrospective study (approval number: 2021–008-01 K) with a waiver of informed consent.

We reviewed consecutive patients with thyroid nodules who had pathology results at our institution between 2017 and 2022. The inclusion criteria included: 1) lesion diameter larger than 4 cm; 2) patients who underwent preoperative thyroid MRI; 3) complete pathology of postoperative specimens. The exclusion criteria included: 1) incomplete clinical and imaging data; 2) poor image quality; 3) lack of contrast enhancement on MRI. Figure 1 displays the study flowchart.

Finally, 82 lesions from 78 patients (32 males and 50 females; age: 50.26 ± 16.20 years; age range: 15–78 years) met the inclusion criteria. Lesions were categorized into benign (n = 62) and malignant (n = 20) groups according to the postoperative pathology.

The 1.5 T MRI scanner (Excite HD; GE Healthcare, Waukesha, WI, USA) used for all MRI examinations was set up with an 8-channel customized neck surface coil (Chenguang Medical Technology Ltd, Shanghai, China). The scan covered the thoracic inlet to the base of the cranium were covered by the scan. The MRI sequences used (CE-T1WI) included axial and coronal fat-suppressed T2-weighted imaging (T2WI), axial T1-weighted imaging (T1WI), single-shot spin-echo echo-planar imaging (SS-SE-EPI) DWI at b values of 0 and 800 s/mm², and axial multiphasic contrast-enhanced T1WI comprised the MRI sequences used (CE-T1WI). A gadolinium contrast agent (Magnevist; Bayer Healthcare, Berlin, Germany) was injected for the CE-T1WI acquisition at a dose of 0.2 mL/kg and a rate of 3 mL/s, followed immediately by 20 mL of physiological saline flushing. Following the injection of the contrast agent, six phases were recorded at intervals of 30, 60, 120, 180, 240, and 300 s intervals while the patients were asked to hold their breath. Table S1 lists detailed acquisition parameters.

ADC maps were automatically created from DWI images (b = 0 and 800 s/mm²) on the console using mono-exponential fitting. Quantitative DWI parameters were measured by two MRI diagnosticians who were blind to the lesion pathology (a chief physician with eight years of experience and a resident with one year each in thyroid MRI diagnosis), using picture archiving and communication system (PACS) and Advantage Workstation 4.5 (GE Healthcare, Waukesha, WI, USA). The section of the whole solid leision portion of the lesion with maximum transverse diameter (excluding cystic, hemorrhage, necrosis, calcium, and vascular structures) was selected to delineate the first region of interest (ROI 1). The following quantitative features in the ROI 1 were measured: 1) mean DWI signal intensity (DWI_SI); 2) mean ADC signal intensity (ADC_SI); 3) mean ADC value (ADC_mean) and minimum ADC value (ADC_min). Another ROI with an 8–10 mm² area is also outlined as a relatively homogeneous solid part without cystic, hemorrhage, necrosis, calcium and vascular structures in the lesion and contralateral to the normal thyroid tissue. The following quantitative features of the ROI 2 were measured in the lesion: 1) DWI signal intensity standard deviation (DWI_SD) and ADC signal intensity standard deviation (ADC_SD) and ADC value standard deviation (ADC_VSD); 2) mean DWI signal intensity and mean ADC signal intensity of contralateral normal thyroid tissue (DWI_NSI and ADC_NSI). The following formulas, DWI_SIR = DWI_SI / DWI_NSI and ADC_SIR = ADC_SI / ADC_NSI, were used to calculate the DWI signal intensity rate (DWI_SIR) and ADC signal intensity rate (ADC_SIR). DWI images (b = 800 s/mm²) and ADC map generated from DWI images (b = 0 and 800 s/mm²) were used for quantitative parameters extraction. All measurements were performed twice and averaged. Figure 2a demonstrates representative images of ROI delineation.

Two US experts retrospectively reviewed US images of thyroid lesions, reaching a consensus without knowledge of the lesion pathology. All lesions with category ≥ 4 were considered malignant according to ACR-TIRADS.

The malignancy prediction model was built using independent factors that were found using univariate and multivariate logistic stepwise regression. By optimizing the Youden's index, receiver operating characteristic (ROC) curve analysis was used to determine the ideal threshold values for the pertinent parameters. Combined thresholds approaches were established based on malignancy-related parameters (Supplementary Method). Individual parameters and models were evaluated using ROC curves, with the area under the ROC curve (AUC) compared by the DeLong test. Unnecessary biopsy rate was defined as the percentage of benign lesions for those requiring biopsy. The diagnostic performance measures for each model, including as accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and unnecessary biopsy rate were calculated, comparing the ACR-TIRADS result was compared.

Table 1 lists the clinicopathological features of thyroid nodules. Except for the location (P = 0.015), there was no difference in the distribution of other features in the benign and malignant thyroid nodules. Table S2 demonstrates the pathological types of thyroid nodules.

Figure 2 shows representative DWI images and the ROI delineation. Moreover, Table 2 demonstrates the results of the univariate and multivariate logistic regression analyses in predicting malignant thyroid nodules. Malignant nodules displayed significantly greater DWI_SD (P = 0.002) and DWI_SIR (P = 0.007) than benign nodules. Additionally, malignant nodules had significantly lower ADC_SD (P = 0.005), ADC_SIR (P = 0.008), ADC_min (P < 0.001), and ADC_mean (P < 0.001) than benign nodules. The ICCs of DWI_SD, DWI_SIR, ADC_SD, ADC_SIR, ADC_min and ADC_mean were 0.776, 0.758, 0.720, 0.923, 0.789, 0.783 and 0.743, respectively. ADC_min was the best-performing parameter with an AUC of 0.933 (0.874—0.992). Figure 3a and Table 3 represent the ROC curves and diagnostic performance metrics at the optimal threshold of relevant individual parameters, respectively.

The optimal threshold values were 1.13 × 10^–3 mm²/s for ADC_min, 1.25 for ADC_SIR, and 1.20 for DWI_SIR, showing their distribution in Fig. 4. For benign and malignant nodules, DWI_SIR, ADC_SIR, and ADC_min overlapped; however, malignant nodule ADC_min was comparatively low.

Table 4 and Fig. 3b depict the diagnostic performance of the multivariate prediction model, combined threshold model, and ACR-TIRADS for malignant and benign thyroid nodules. The multivariate prediction model had the best diagnostic performance with an AUC of 0.946 (0.896–0.996) at a cutoff value of 0.198, which was higher than the AUC achieved by the combined threshold model (DWI_SIR and ADC_min), with insignificance difference (P = 0.500). The AUC of combined threshold model (DWI_SIR and ADC _SIR) and ACR-TIRADS were 0.777 (0.648–0.907) and 0.722 (0.588–0.857), respectively. Figure 5 reveals the grouped scatter plots of the two combined threshold models, and Table S3 summarizes the Delong test results for AUC comparison among different models.

The sensitivity (90.0%) and NPV (96.6%) were the highest in the multivariate prediction model. The results showed three false negative lesions, all follicular thyroid carcinoma, and seven false positive lesions: three nodular goiters, three adenomatous nodular goiters, and one adenoma. The best specificity and PPV (both 100%) were achieved by the combined threshold model (DWI_SIR and ADC_min), where five false negative lesions were all follicular thyroid carcinomas. The combined DWI_SIR and ADC_min had the lowest unnecessary biopsy rate with no false positive cases. The accuracy, sensitivity, specificity, PPV and NPV of the combined thresholds model (DWI_SIR and ADC_SIR) were 81.7%, 70%, 85.5%, 60.9% and 89.2%, respectively. Compared with ACR-TIRADS, the quantitative DWI parameter-based models significantly improved differentiating benign and malignant thyroid nodules.