This Week in Radiology — Aug 15, 2026

Welcome to This Week in Radiology. This week we're covering 10 notable papers spanning artificial intelligence in the reading room and the reporting pipeline, breast imaging from density scoring to biopsy markers, and advances in neuro and musculoskeletal imaging hardware and stroke care. Let's dive in.

We start with artificial intelligence, where the interesting question has shifted from whether algorithms can detect disease to what happens when you actually plug them into a working department. In European Radiology, a prospective alternating-month study switched a commercial mammography algorithm on and off, month by month, across more than four and a half thousand examinations read by four radiologists [1]. During the assisted months, the cancer detection rate roughly doubled, from about eleven and a half to about twenty-two per thousand examinations, and reading time did not budge at all, sitting at around sixty-five seconds either way. That is reassuring for anyone worried that decision support adds friction. The important nuance is where the false-positive cost landed. For screening examinations, the abnormal interpretation rate was statistically unchanged at around nine percent. For diagnostic examinations, though, abnormal interpretations rose from about twelve to nearly nineteen percent — a meaningful jump in recalls and further workup in a population that already has a clinical problem. The practical message is that the same tool behaves differently depending on the indication, and diagnostic mammography deserves more caution than screening.

Staying in European Radiology and in breast screening, a second paper examined automated volumetric density scoring from a commercial artificial intelligence model in two hundred thousand examinations from BreastScreen Norway [2]. Cancer detection performance was strong overall, but it fell steadily as breasts got denser — the area under the curve dropped from about 0.96 in the fattiest category to about 0.86 in the densest. Two findings should give pause to anyone planning density-based risk-stratified screening. First, the proportion of women labelled as extremely dense differed by mammography vendor, roughly three percent with one manufacturer versus nearly seven percent with the other. Second, at the individual breast level, close to one in five examinations assigned the right and left breast to different density categories. So the same woman could be classified differently depending on which machine she attended and whether you score the examination or the breast. Read alongside the reading-time study, the theme is consistent: these tools work, but their behaviour is context-dependent, and local validation is not optional.

Artificial intelligence also showed up this week in the unglamorous but high-yield space of reporting. In the Journal of the American College of Radiology, a group implemented a vendor-agnostic optical character recognition pipeline that reads numbers straight off bone density images and drafts complete reports [4]. Across four sites, report creation time fell from about three and a half minutes to under a minute at the academic sites and from about one and a half minutes to under forty seconds in the community. The turnaround figures are the striking part: community-site turnaround dropped from roughly one hundred thirty-four hours to forty-two hours, about a four-day improvement. Accuracy was preserved and in fact marginally better than the original human reports, and completeness at community sites went from forty-five percent to one hundred percent. This is a reminder that some of the largest available gains are in structured, numeric, template-driven reporting rather than in diagnosis itself.

Our second theme is breast intervention and the tissue-imaging mismatch problem. In the European Journal of Radiology, a single-centre series of three hundred thirty-four patients used ultrasound-guided vacuum-assisted excision with a nine-gauge device to work up lesions where the imaging appearance and the core needle histology disagreed [5]. After excision, about seventy percent of lesions were benign, one in five were B3 lesions, and just over nine percent were upgraded to malignancy. Delayed false negatives were about one percent, specificity and positive predictive value were one hundred percent, and there were no major complications. BI-RADS category, morphology, margins and orientation all predicted upgrade. The practice implication is that for discordant lesions, a percutaneous vacuum excision can resolve the discordance in most women and spare them a diagnostic operation, provided you accept a residual upgrade rate around one in ten and follow the rest. Complementing that, RadioGraphics offers a review of biopsy marker selection — the clips we place and largely stop thinking about [9]. It traces markers from simple metal to multimaterial devices that now serve cross-modality correlation, surgical planning and radiation targeting, with good long-term safety and cost-effectiveness through fewer re-excisions. The contemporary problems are marker migration, allergic reactions to metallic or embedding components, and poor ultrasound visibility — addressed by nonmetallic designs, careful placement timing, migration-prevention technique, and exploiting the Doppler twinkling artifact to find the clip on ultrasound.

Our third theme is hardware and technique, where two papers ask whether new detectors and new software actually change what we see. In the American Journal of Neuroradiology, a pragmatic paired comparison put clinical photon-counting CT against conventional energy-integrating CT in one hundred ninety-four patients contributing two hundred seventy matched examinations, read blinded and at least four weeks apart [3]. In the main non-contrast brain cohort of two hundred twenty-five pairs, photon-counting scored higher on overall image quality, and pathology conspicuity favoured it in essentially every pair — two hundred twenty-four of two hundred twenty-five. Grey-to-white matter contrast-to-noise roughly doubled, dose-normalised contrast-to-noise rose even more, and median dose fell about twenty-two percent, from about forty-nine to thirty-eight milligray. Artifact scores were no different. The authors are appropriately candid that scanner generation, vendor, protocol and reconstruction were all unmatched, so this is a real-world system comparison rather than proof that the detector alone is responsible. Pointing in a similar direction from the software side, the European Journal of Radiology reports thirty-three knee examinations in twenty-six patients scanned at both 0.55 tesla and 3 tesla, with deep learning denoising applied to the low-field data [7]. Unprocessed 0.55 tesla imaging achieved sensitivity and accuracy around 0.83. After denoising, sensitivity rose to 0.97 and accuracy to 0.98, with excellent agreement against 3 tesla. Image quality and reader confidence improved with denoising but still remained below 3 tesla. In a small, retrospective, single-centre cohort, that is encouraging for low-field musculoskeletal imaging in resource-limited or point-of-care settings, without yet displacing high-field magnets.

Our final theme is neurologic disease, and here the notable results are the negative ones. In the American Journal of Neuroradiology, investigators asked whether baseline amyloid burden on PET, expressed as a Centiloid score, predicts amyloid-related imaging abnormalities in patients on lecanemab [6]. Across forty-one patients with these abnormalities and seventy-two controls, and after adjusting for age, sex, APOE epsilon-4 status and baseline microbleeds, baseline Centiloid was not associated with the complication overall, nor with the haemorrhagic or oedematous subtypes, nor with time to onset. An exploratory regional analysis did find modestly higher baseline uptake in the specific regions that later developed oedema compared with mirrored contralateral regions, which is hypothesis-generating only. So global amyloid quantification should not currently be used to triage which patients are safe to treat. Also in the same journal, a propensity-matched comparison of tenecteplase versus alteplase before thrombectomy in patients aged eighty and older analysed two hundred seventy-eight matched pairs [8]. Good functional outcome at ninety days was similar — forty-eight versus forty-three and a half percent, with confidence intervals crossing one — as were early reperfusion, symptomatic haemorrhage, and ninety-day mortality at thirty percent in both. Complete neurological recovery at twenty-four hours was more frequent with tenecteplase, but the authors show that signal was confined to a single scale value, left no trace at ninety days, and is vulnerable to unmeasured confounding since drug was completely confounded with centre. The honest conclusion is equivalence, and randomised data in the very elderly are still needed. Rounding out the neuro section, RadioGraphics provides a review of lacunar stroke syndromes, which make up about a quarter of all strokes, covering the five classic syndromes, atypical presentations, and the mimics that trip readers up when infarcts are too small to see confidently [10].

If you only have time for one paper this week, make it the alternating-month mammography study in European Radiology [1]. It is one of the few prospective, in-workflow tests of artificial intelligence assistance that measures both the benefit and the cost, and its indication-specific finding should shape how your department deploys these tools.

Here are the key takeaways from this week in Radiology. Artificial intelligence assistance in routine mammography raised cancer detection without slowing readers down, but increased abnormal interpretations in diagnostic examinations only — deploy it deliberately by indication. Automated density scoring varies by mammography vendor and between the two breasts of the same woman, so validate locally before building risk-stratified screening on it. Some of the biggest operational wins from artificial intelligence are in structured numeric reporting, where automated bone density drafting cut community turnaround by about four days while improving completeness. Ultrasound-guided vacuum-assisted excision resolves imaging-histology discordance in most patients and can spare surgery, with an upgrade rate around one in ten. Photon-counting CT delivered better brain image quality and conspicuity at roughly a fifth lower dose in real-world use, and deep learning denoising brought low-field knee MRI close to 3 tesla accuracy. And two negative results matter: baseline amyloid Centiloid does not predict amyloid-related imaging abnormalities on lecanemab, and tenecteplase and alteplase performed comparably before thrombectomy in patients over eighty.

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 if you have a paper you have been meaning to read, upload the PDF, or paste any link, at audioscholar dot C C. We will turn it into audio like this one, in any of thirty-one languages.

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