The fractures radiographs hide are the ones a model found

The edition in brief

Today's radiology desk opens with OccuNet, a deep learning model tested across four hospitals in 2,576 adults with suspected hip trauma, all of whom had radiographs plus same-episode CT or MRI. On the pooled test set the model reached 97.5% sensitivity and 98.8% specificity, area under the curve 0.99. In the 189 radiograph-negative or indeterminate fractures it reached 94.7% sensitivity against 86.2% for musculoskeletal radiologists and 68.8% for emergency physicians; with the model as an aid, reader sensitivity rose and reading time fell by about 15 to 19%. The American College of Radiology has updated its appropriateness criteria for myelopathy, covering acute, chronic and vascular causes. An analysis of 904,209 head CT scans from an international dose registry found size-adjusted dose-length product varying more than twofold, and up to 3.4-fold, between clusters of commonly used protocols — for adult helical scans, 372 to 998 mGy·cm. Photon-counting CT for coronary stents cut non-diagnostic stent assessments from 46.8% to 1.1% and reached 88.5% per-stent accuracy for in-stent restenosis of 50% or more, with a negative predictive value of 100% but positive predictive value of only 56.3%. A pearl addresses the persistent myth linking shellfish allergy to iodinated contrast. The edition closes with a randomised trial in which radiologist-administered ketamine plus midazolam beat fentanyl plus midazolam on pain, with fewer desaturations.