Performance and Reproducibility of Bone-RADS MRI Algorithm for Vertebral Lesions
Emil J Barkovich1, William A Mehan2, Karen Buch2
1From the Department of Radiology and Biomedical Imaging (E.J.B.), University of California, San Francisco, 505 Parnassus Ave, San Francisco, CA 94143 and Department of Radiology (W.A.M., K.B.), Massachusetts General Hospital, Harvard Medical School, 55 Fruit St, Boston, MA 02114. embarko@comcast.net.
Purpose:
To evaluate the inter-rater reliability and accuracy of the SSR Bone-RADS MRI algorithm for challenging osseous vertebral lesions.
Materials And Methods:
We retrospectively reviewed 538 MRI exams with vertebral lesions described in the radiology reports resulting from a keyword search of "indeterminate." Studies were excluded if protocols lacked T1WI or T2WI or if there was known history of vertebral metastases. Based on PACS and EMR review, lesions were categorized as stable, incomplete follow-up, typically benign or degenerative, malignant or no follow-up. Per Bone-RADS, stability required follow-up at ≥24 months. No change at 2-24 months was considered incomplete follow-up and imaging at <2 months was considered no follow-up. Lesions were also defined as stable if CT imaging >2 years later showed no change or normal appearing bone. Malignancy status was divided into (1) none known, (2) no known metastatic disease and (3) known metastatic disease. Two CAQ-certified neuroradiologists independently reviewed the images and scored each lesion using Bone-RADS. They did not incorporate Bone-RADS clinical criteria and cases with multiple lesions were not excluded. Discrepant scores were reviewed for consensus. Predictive values and inter-rater reliability were calculated.
Results:
282 vertebral lesions met inclusion criteria. 60% had subsequent diagnosis or imaging follow-up of ≥2 months. 66% of these patients had no reported malignancy history. There were four cases of newly diagnosed bone metastases and two cases of plasmacytoma. The negative predictive value (NPV) of Bone-RADS 1 was 100% (95% CI [94.7-100%], n = 46). Positive predictive value for indeterminate lesions was 3.4% (95% CI [1.2-8.0%], n = 116); when Bone-RADS clinical criteria were applied this decreased to 0% (95% CI [0-3.0%], n = 81). Cohen's kappa test for inter-reader reliability was 0.60. Of cases scored as Bone-RADS 1, 80% of original radiologist reports had advised follow-up.
Conclusion:
A Bone-RADS 1 score showed excellent NPV for vertebral lesions. Lesions scored as 2/3 had low malignancy rate which decreased to 0% when exclusionary clinical criteria were incorporated. The Bone-RADS MRI algorithm had moderate inter-reader reliability on our cohort. Given intermediate inter-reader reliability and the high prevalence of both benign and metastatic vertebral lesions relative to primary bone tumors, radiologists should be cautious when applying Bone-RADS to this anatomic region. A vertebral specific algorithm would be helpful.
