The edition · Radiology
The model ranked the applicants, and pushed the same groups down every time
Seven model configurations produced 140 rank order lists from application data alone. None came close to the committee's, and all of them placed women, international and non-MD applicants lower than the humans did. Plus what a skull-base report has to contain, where robotic CT guidance earns its place, and a national registry using language models to watch imaging algorithms.
The edition in brief
Today's radiology desk opens with imaging for endoscopic endonasal skull-base surgery. With the region effectively invisible on examination, the operative plan is built from high-resolution CT and MRI, and the review sets out the normal and variant anatomy, the surgical landmarks and the cautionary findings an interpreting radiologist has to name if the surgeon is to choose the right corridor and plan reconstruction. A narrative review of robot-assisted CT-guided intervention, mapped onto the IDEAL framework, finds preclinical work consistently showing better targeting and fewer needle adjustments, clinical evidence concentrated in liver, lung, kidney and musculoskeletal procedures with long, oblique or out-of-plane trajectories, and operator radiation exposure frequently reduced - while procedure time and patient dose vary and most studies sit at IDEAL stages 1 to 2b. The American College of Radiology's Assess-AI registry reports language-model prompts for nine use cases extracting findings from reports, with agreement against report-derived reference labels of 0.985 for intracranial haemorrhage and 0.997 for pulmonary embolism - figures the authors are careful to say came from cohorts that also shaped the prompts and the labels, and so are not independent validation. The edition closes on residency selection. Across 148 applicants and 140 generated lists, models working from pre-interview application data alone agreed poorly with the committee's final rank order (median Kendall tau 0.15 to 0.36), while adding interviewer scores raised agreement to 0.84 to 0.93 and sorting by interview score alone reached 0.83. In the application-only condition the models systematically placed female, international and non-MD applicants below the committee's positions, and programme signallers above them.
What the skull-base surgeon actually needs from your report
Report skull-base studies against the operative corridor, naming the structures at risk and the reconstruction implications, rather than describing the lesion alone.
Robotic CT guidance: the case is strongest where the trajectory is hardest
Consider robotic CT guidance for the technically difficult trajectory, and do not expect it to pay back on routine work.
A national registry using language models to watch imaging algorithms
Ask who is monitoring the imaging algorithms running in your department, and against what reference - the answer is often nobody and nothing.
Write down when you disagreed with the algorithm
Log every disagreement between the algorithm and your read - without it, drift is undetectable.
Language models could not reproduce a rank order list, and displaced the same groups doing it
Keep language models out of rank order list generation; use them for retrieval and audit, and check any ranking tool for subgroup displacement before it goes near a decision.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this one is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for radiology, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free