The accuracy of thyroid nodule ultrasound to predict thyroid cancer: systematic review and meta-analysis

Abstract

Context: Significant uncertainty remains surrounding the diagnostic accuracy of sonographic features used to predict the malignant potential of thyroid nodules.

Objective: The objective of the study was to summarize the available literature related to the accuracy of thyroid nodule ultrasound (US) in the prediction of thyroid cancer.

Methods: We searched multiple databases and reference lists for cohort studies that enrolled adults with thyroid nodules with reported diagnostic measures of sonography. A total of 14 relevant US features were analyzed.

Results: We included 31 studies between 1985 and 2012 (number of nodules studied 18,288; average size 15 mm). The frequency of thyroid cancer was 20%. The most common type of cancer was papillary thyroid cancer (84%). The US nodule features with the highest diagnostic odds ratio for malignancy was being taller than wider [11.14 (95% confidence interval 6.6-18.9)]. Conversely, the US nodule features with the highest diagnostic odds ratio for benign nodules was spongiform appearance [12 (95% confidence interval 0.61-234.3)]. Heterogeneity across studies was substantial. Estimates of accuracy depended on the experience of the physician interpreting the US, the type of cancer and nodule (indeterminate), and type of reference standard. In a threshold model, spongiform appearance and cystic nodules were the only two features that, if present, could have avoided the use of fine-needle aspiration biopsy.

Conclusions: Low- to moderate-quality evidence suggests that individual ultrasound features are not accurate predictors of thyroid cancer. Two features, cystic content and spongiform appearance, however, might predict benign nodules, but this has limited applicability to clinical practice due to their infrequent occurrence.

Figure 1.

Study selection.

Figure 2.

Quality assessment of included studies with the four main domains of QUADAS2 to evaluate the risk of bias: 1) selection of patients: could the selection of patients have introduced bias?; 2) index test: could the conduct or interpretation of the index test have introduced bias?; 3) standard test: could the reference standard, its conduct, or its interpretation have introduced bias?; and 4) flow: could the patient flow have introduced bias?.

Figure 3.

Fagan nomogram representing the LR for positive results (blue) and the LR for negative results (red) of the four most notable features. The nomogram has three components: 1) pretest probability, which is the estimated prevalence of thyroid cancer (overall prevalence of thyroid cancer in this review 20%); 2) LR of the feature; 3) posttest probability, which is the probability of having the condition given that feature is present (blue) or absent (red). The LR for spongiform and cystic feature was calculated based on the LR to predict benign nodule (eg, the false positives for benign nodules is the true positive for malignant nodule).