Illustrative AI review sample — radiology / imaging AI

Key Points

1Single-centre Cambridge cohort of 370 unscreened prostate cancer patients evaluating MRI integration into pre-treatment risk stratification; clinically meaningful reclassification rate shown, but retrospective MRI ascertainment, absent PI-RADS scoring, and limited model calibration reporting weaken the prognostic contribution.

Major Issues

Methods: MRI reporting was retrospective without standardised PI-RADS version 2.1 scoring across the study period; heterogeneous reporting criteria from different radiologists over multiple years introduce significant measurement variability in the MRI-derived predictors.

Results: Prognostic model performance improvement is reported with discrimination statistics (AUC) but calibration statistics (Hosmer–Lemeshow test, calibration slope, calibration plots) are absent; a model with improved discrimination may still be poorly calibrated for individual risk prediction.

Methods / Results: Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) — the standard statistics for prognostic reclassification studies — are not reported; the clinical utility of the 6.2% reclassification rate cannot be formally evaluated without these metrics.

Design: Single-centre design at a tertiary academic centre; patients referred to this level of care are likely more complex and higher-risk than the general unscreened population, limiting generalisability to district general hospital prostate cancer pathways.

Assessing the impact of MRI based diagnostics on pre-treatment disease classification and prognostic model performance in men diagnosed with new prostate cancer from an unscreened population

Key Points

Major Issues