A threefold dose spread on the same head CT, and an AI that finds the fracture you cannot see

The edition in brief

Four findings with direct consequences for how a department scans and reports. Cluster analysis of 864,182 adult and 40,027 paediatric head CT scans from an international dose registry sorted routine protocols into clusters and found size-adjusted dose-length product varying more than two-fold, and up to 3.4-fold, within each stratum - for adult helical scans, 372 to 998 mGy·cm. The lowest-dose clusters generally used lower tube current and voltage, but several protocol routes reached the same place. In a randomised trial of 264 image-guided lung biopsies, bone biopsies and abscess drainages, radiologist-administered ketamine with midazolam reduced maximum intraprocedural pain against fentanyl with midazolam (mean difference -1.4 points, 95% CI -2.0 to -0.8), cut pain scores above 4 from 17% to 2.3%, and produced fewer desaturations below 90% (2.3% against 9.8%), at the cost of more hallucinations (11.4% against 3.8%). A review of the top 5% of AI risk scores in BreastScreen Norway found that in mammograms with no cancer over six years, calcifications alone were the feature the AI marked in about 70% of cases, and 76% had been read as benign; among screen-detected cancers, a spiculated mass was most common. An error-annotation study of 185 patients found process-related diagnostic errors in 14.6% (95% CI 10.2-20.4), most often delay in performing an ordered test. The edition closes on OccuNet, a deep learning model tested across four hospitals and 2,576 patients, which reached 94.7% sensitivity for radiograph-negative or indeterminate femoral neck fractures against 86.2% for radiologists and 68.8% for emergency physicians, and raised reader sensitivity while shortening reading time.