Near-perfect AI results on prostate MRI may reflect study design, not better algorithms

The edition in brief

A perspective in European Radiology appraises artificial intelligence studies for detecting clinically significant prostate cancer on MRI across six domains and concludes that apparent near-perfect performance can arise from preventable design choices such as spectrum bias, pathway bias, imperfect reference standards and confounded comparisons. A meta-analysis of 36 studies and 8,364 patients on MRI pelvimetry before rectal cancer surgery found, in six studies with 769 patients, that shorter interspinous and intertuberous distances were associated with surrogate measures of surgical difficulty, though definitions varied and risk of bias concerns were common. A prospective study of 348 patients and 31 volunteers found higher signal-to-noise ratio and diagnostic confidence at 5.0 T than at 1.5 T and 3.0 T for upper abdominal MRI, with more cardiac motion artefact, signal inhomogeneity on T2 and distortion on diffusion imaging.