Purpose: To evaluate accuracy in interpretation of shoulder magnetic resonance (MR) images and interobserver agreement and to characterize the types of errors.
Materials and methods: Five radiologists with varying experience independently and retrospectively twice interpreted the MR images of 222 symptomatic patients who underwent both MR imaging and shoulder arthroscopy. The first interpretation was a blind review; the second was with knowledge of the arthroscopic findings.
Results: For full-thickness tears, the sensitivity, specificity, and accuracy were 84%-96%, 94%-98%, and 92%-97%, respectively, and for partial tears, 35%-44%, 85%-97%, and 77%-87%, respectively. There was no statistically significant difference between readers in the detection of partial or full-thickness tears. There was a statistically significant difference between readers in the no-tear category. Cohen kappa values generally indicated improved interobserver agreement proportional to the readers' experience (full-thickness tears, 0.731-0.881; partial tears, 0.168-0.443).
Conclusion: Diagnosis of a full-thickness tear can be learned to a high degree of accuracy. Despite the radiologist's level of experience and knowledge of the arthroscopic findings, the sensitivity for diagnosis of partial tears is poor.