Fifteen studies met the inclusion criteria. Eleven studies reported on a per-lesion basis, two on a per-patient basis, and two reported both. Of the 13 studies reporting results on a per-lesion basis (1,802 patients, range 61 to 332), all recruited a representative patient spectrum, used an appropriate reference standard, avoided partial verification bias, blinded interpreters of the index and reference standard tests, and explained withdrawals. Seven studies avoided progression bias, 12 avoided incorporation bias, four avoided differential verification bias, and three reported uninterpretable results. Only one study was prospective.
Based on the per-lesion analysis (13 studies), the summary estimate of diagnostic sensitivity was 93% (95% CI 90 to 95; Ι²=61%), diagnostic specificity was 95% (95% CI 91 to 97; Ι²=57%), positive likelihood ratio was 18.07 (95% CI 10.52 to 31.04; Ι²=14%), negative likelihood ratio was 0.07 (95% CI 0.05 to 0.10; Ι²=62%), the diagnostic odds ratio was 249.81 (95% CI 125.12 to 498.74), and the area under the curve was 0.98 (95% CI 0.96 to 0.99).
Meta-regression showed that the reference standard, study nation, publish year, study design, MRI field strength, or location of primary tumour were not strongly associated with accuracy. A threshold effect was not observed. Sensitivity was significantly lower for lesions under 10 mm than for lesions over 10 mm.
Further results from sensitivity analyses were reported.
The potential for notable publication bias was observed.