The model helped the residents most, and misled them most

The edition in brief

Today's radiology edition opens with a reader study of 93 thoracic cases in which ten readers — five thoracic radiologists and five residents — described the findings in free text, fed that description to a large language model, and then chose again. Accuracy rose in both groups, from 56.3% to 65.6% for radiologists and from 42.4% to 58.5% for residents. The residents gained more, but they also accepted the model's preferred diagnosis far more often (73.2% vs 48.9%) and switched to a wrong answer after a misleading output in 60.6% of such cases against 32.4%. The model was also better when a specialist wrote the description than when a resident did. An AJR expert panel sets out why paediatric imaging is so poorly served by artificial intelligence — scarce datasets, developmental variation, off-label use of adult-trained models, absent post-deployment surveillance — and what a paediatric implementation roadmap needs. A deep learning emphysema segmentation outperformed the conventional LAA-950 threshold on Dice accuracy, bias and correlation with visual scores and lung function across three clinical cohorts including 1,159 patients. A prospective study of 34 patients with head and neck squamous cell carcinoma found a PD-L1-targeted tracer, 68Ga-DOTA-WL12, discriminated PD-L1 positivity far better than FDG (area under the curve 0.92 vs 0.69) and independently predicted chemo-immunotherapy response. The practice-changer is carotid Plaque-RADS, where adding plaque burden, remodelling index, enhancement ratio and perivascular fat density raised discrimination of symptomatic plaque to 0.919 in the moderate-stenosis group that stenosis measurement alone cannot resolve.