This Week in Radiology — Sep 19, 2026

Welcome to This Week in Radiology. This week we're covering 10 notable papers spanning artificial intelligence moving from proof of concept into the reading room, screening and early detection strategies, oncologic staging with PET, and some uncomfortable data on who actually gets the imaging guidelines say they should. Let's dive in.

We'll start with artificial intelligence, because the most striking paper of the week comes from Science, where Zhang and colleagues describe RADAR, a generalist vision-language model for abdominal CT [1]. Rather than training a narrow classifier on one disease, they trained on more than four hundred thousand contrast-enhanced abdominal CT examinations and roughly fifteen million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation. Across internal and external evaluation at multiple centres, the model covered eighteen anatomical structures and one hundred and forty-six imaging findings, and in a reader study involving twenty-six radiologists, working with the model raised diagnostic sensitivity by about ten percent. The significance here is the scope: this is not a pulmonary nodule detector, it is a system attempting the breadth of a general abdominal read, and it performed at expert level on both routine and complicated tasks. Whether that generalises outside the training environment is the open question, but the direction of travel is clear. A narrower but instructive example comes from the European Journal of Radiology, where Yu and colleagues built a multimodal deep learning model for acute pulmonary embolism using spectral CT pulmonary angiography [2]. They studied five hundred and twenty-six participants across three cohorts, of whom one hundred and twenty-six had pulmonary embolism, and fused the conventional angiographic images with iodine density maps. Be careful with how you read the results: at the patient level, the multimodal model did not significantly outperform conventional angiography alone. Where it did help, and help meaningfully, was in slice-level detection, lesion segmentation, and particularly peripheral emboli, where recall rose from roughly half of peripheral clots to just under two thirds internally, and from just under sixty percent to just over seventy percent in the external test set. This is a pilot study, and the authors say so plainly.

Against those technical papers, the American Journal of Roentgenology offers something quite different: an expert panel opinion from Petrella and colleagues on artificial intelligence detection of amyloid-related imaging abnormalities during anti-amyloid therapy for Alzheimer disease [3]. This is a real and growing workload problem. Surveillance MRI volumes are climbing, interreader variability for subtle findings is substantial, and a missed abnormality changes dosing and patient safety. The multidisciplinary panel of neuroradiologists and Alzheimer clinicians concluded that these tools are likely to improve safety, but only as clinical decision support within a radiologist-in-the-loop framework. They flagged substantial variation among commercially available products in regulatory status, technical capability, and the validation evidence behind them, and they noted that no study has yet shown that better detection translates into better patient outcomes. Their recommendation was conditional implementation with radiologist oversight, ongoing quality assurance, and prospective performance monitoring. If your department is being sold one of these tools, that list is your due diligence checklist.

Our second theme is finding disease before it declares itself, and RadioGraphics contributes two reviews that pull in opposite emotional directions. Dogra and colleagues review opportunistic CT screening, the extraction of biomarkers like low bone mineral density, hepatic steatosis, sarcopenia, adiposity, and vascular calcification from scans done for entirely unrelated reasons [4]. The appeal is obvious: no extra radiation, no extra scanner time, no extra patient burden, and automated analysis now makes it feasible at scale. The caution the authors deliver is equally important, and it's the part worth carrying into practice: not all of these biomarkers are equally actionable, and indiscriminate reporting risks overdiagnosis and unnecessary follow-up. They spend as much time on workflow integration, measurement standardisation, and post-deployment algorithm monitoring as on the biomarkers themselves. Srinivas Rao and colleagues, also in RadioGraphics, review where pancreatic cancer screening actually stands [5]. Five-year survival remains around thirteen percent, driven almost entirely by late diagnosis, and screening is currently justified only in genuinely high-risk individuals, those with a relevant family history, predisposing genes or inherited syndromes, or known precursor lesions. Endoscopic ultrasound, MRI, and CT are the supported primary modalities, with intervals that vary by risk category. Their honest assessment is that conventional imaging still misses visually occult disease, and that artificial intelligence applied to prediagnostic CT and MRI, alongside more sensitive serum biomarkers, is where the field is heading, but that validated, cost-effective, scalable tools do not yet exist.

Moving to oncologic staging and response assessment, European Radiology published two papers that should influence how you advise your multidisciplinary teams. Moore and colleagues compared FDG PET-CT with contrast-enhanced CT after neoadjuvant chemotherapy for oesophagogastric adenocarcinoma, in sixty-six patients who had both studies performed contemporaneously on the same visit [6]. For predicting pathological tumour stage and nodal stage, the two modalities were essentially equivalent, and frankly both were mediocre, with nodal staging performing at close to chance. Where PET separated itself was in predicting pathological treatment response, where it was clearly better than CT, which performed no better than a coin toss. Eight patients had metastatic disease on restaging; PET identified all eight, CT identified four, though with those small numbers that difference did not reach statistical significance. The practical reading is that PET earns its cost in advanced disease where occult metastases are likely and where response assessment will actually change management. In the same journal, Kleiburg and colleagues studied two hundred and four patients with synchronous metastatic hormone-sensitive prostate cancer staged by PSMA PET/CT [7], asking whether our inherited low-volume and high-volume categories still make sense. Multidisciplinary team classification did work: median overall survival was sixty-nine months in low-volume disease versus forty-one months in high-volume disease. But quantitative PSMA total tumour volume was an independent predictor of survival, with each doubling of tumour volume associated with roughly a forty percent increase in mortality risk, independent of performance status and treatment. Their exploratory threshold, a PSMA tumour volume of one hundred and fifty millilitres or any visceral metastasis, performed comparably to expert consensus. Their more interesting argument is that we should stop dichotomising altogether and use tumour volume as a continuous variable. Multicentre validation is needed before anyone changes practice. Alongside these, European Radiology also published a review from Martín-Noguerol and colleagues on photon-counting CT in head and neck imaging [8], making the case that higher spatial resolution, lower noise, reduced metal artifact, and spectral reconstructions act synergistically, which matters most in patients with cochlear implants, dental hardware or postoperative material, and in children and those under repeated surveillance where dose and contrast reduction accumulate.

Finally, two papers from the Journal of the American College of Radiology on how care is actually delivered. Mattke and colleagues examined structural brain imaging after a new diagnosis of cognitive impairment, using complete national Medicare and Medicare Advantage data covering nearly one point one million newly diagnosed beneficiaries aged sixty-five to eighty-four [9]. Guidelines recommend structural imaging, yet only about fifty-seven percent of patients received an MRI or CT within a year either side of diagnosis, about sixty-six percent for dementia but only about half of those with mild cognitive impairment. Rates were lower in Medicare Advantage members and lower again for dually eligible beneficiaries, and MRI was the initial study in fewer than a third of cases, with the lowest MRI use in exactly those same disadvantaged groups. In an era where anti-amyloid therapy eligibility depends on baseline imaging, this is a substantial and inequitable gap. Also in that journal, Wilkins and colleagues analysed nearly thirty thousand practice-years of Medicare data on interventional radiology [10] and found that the degree of an individual radiologist's specialisation in interventional work, far more than the practice's overall interventional share, predicted the complexity of procedures performed, and that evaluation and management services, including those delivered by non-physician practitioners, were strongly associated with higher complexity work. The message for practice leaders is that clinical service infrastructure and individual subspecialisation are what sustain complex interventional care.

If you only have time for one paper this week, make it the RADAR generalist model in Science [1]. It is the first credible demonstration that a single artificial intelligence system can cover the breadth of abdominal CT rather than one narrow task, and the reader study showing improved radiologist sensitivity tells you this is arriving as an assistive tool, not a hypothetical one.

Here are the key takeaways from this week in Radiology. Generalist artificial intelligence for abdominal CT has reached expert-level performance and improved radiologists' sensitivity in a reader study, while a spectral CT pulmonary angiography model helped mainly with peripheral emboli and not with patient-level detection. When evaluating any commercial tool, including the amyloid-related imaging abnormality detectors, insist on regulatory clarity, local validation, radiologist oversight, and prospective monitoring. Opportunistic CT screening is technically ready but should be deployed selectively, favouring actionable biomarkers over indiscriminate reporting. In oesophagogastric cancer, PET adds value for treatment response and occult metastases rather than for T and N staging, and in metastatic hormone-sensitive prostate cancer, quantitative PSMA tumour volume improves prognostic stratification beyond our current volume categories. And nearly half of Medicare patients with a new cognitive impairment diagnosis are not getting guideline-recommended structural imaging, with the largest gaps among the most disadvantaged.

That's your roundup for This Week in Radiology. The full transcript and references are available on the episode page in your AudioScholar library. This is an AI-curated summary — for clinical decisions, always consult primary sources and current guidelines. See you next week.

If this weekly briefing is useful, follow the show in your podcast app so new episodes arrive automatically. And a quick rating on Apple Podcasts or Spotify helps other physicians discover the show.

This is an automated summary generated by artificial intelligence, which can make mistakes. Always review the original source materials.