This Week in Radiology — Sep 12, 2026

Welcome to This Week in Radiology. This week we're covering 10 notable papers spanning three broad themes: screening imaging in special populations, advanced quantitative and artificial intelligence-driven image analysis, and imaging that guides or follows interventional and oncologic therapy. Let's dive in.

We start with screening in populations where the usual rules don't apply. RadioGraphics offers a practical review from Fazeli and colleagues on breast cancer screening during pregnancy and lactation [1]. The context matters: childbearing is being delayed, and risk assessment now often begins at age 25, so more patients are pregnant or lactating when their high-risk screening is due. The key operational points are that mammography is safe in pregnancy and lactation but less sensitive, because hormonal changes increase parenchymal density and can mask malignancy; ultrasound avoids ionising radiation and remains the primary tool for symptomatic patients, though its value as a routine screening test in this group is genuinely unproven; and MRI is contraindicated in pregnancy but can be used during lactation for high-risk screening or newly diagnosed disease, accepting marked background parenchymal enhancement. The practical takeaway is to resist the reflex to defer imaging. Pregnancy-associated breast cancer is rising in incidence and delayed diagnosis is the recurring failure mode, so a palpable lump in a lactating patient deserves the same urgency it would get in anyone else. A companion RadioGraphics review by El Omeiri and colleagues applies the same risk-based logic at the other end of life, summarising targeted ultrasound screening in high-risk newborns and infants, including preterm infants, breech presentations, congenital and syndromic conditions, and tumour predisposition syndromes [2]. Their unifying principle is worth repeating in any department: risk-based selection rather than universal imaging, standardised protocols, and interpretation by radiologists who know the population-specific disease patterns.

The second theme is quantification — where algorithms and structured reporting are being asked to outperform the single threshold or the single number we have relied on for years. In European Radiology, Sotoudeh-Paima and colleagues built a deep learning segmentation model for emphysema on chest CT, pre-trained on virtual patient data across scanners, doses and kernels, then fine-tuned on expert-annotated masks, and tested it against the conventional low-attenuation-area threshold of minus 950 Hounsfield units [5]. Across virtual and clinical cohorts, including more than eleven hundred patients in the largest cohort, segmentation accuracy was roughly doubled compared with the threshold method, agreement across imaging conditions was tighter, and correlation with radiologist visual scoring rose from weak to strong. It also tracked pulmonary function, particularly diffusing capacity, better than the threshold biomarker. If your practice reports quantitative emphysema, this is a signal that the venerable minus 950 cut-off is nearing the end of its useful life. In the same journal, Luo and colleagues took the carotid Plaque-RADS system and asked whether ancillary features add anything [3]. In a retrospective cohort imaged with both CT angiography and high-resolution vessel wall MRI, plaque burden, remodelling index, enhancement ratio and perivascular fat density were each independently associated with symptomatic plaque. An integrated model combining Plaque-RADS, stenosis degree and those ancillary features discriminated symptomatic from asymptomatic plaque well overall, and performed even better in the moderate stenosis subgroup — precisely the patients where stenosis alone leaves you undecided. Also in European Journal of Radiology, García-Hidalgo and colleagues challenge the dominance of cerebral blood volume in dynamic susceptibility contrast perfusion imaging of brain tumours [9]. In 219 presurgical patients scanned twice in a single session with different acquisition parameters, adding microvascular metrics such as capillary transit time heterogeneity and metabolic metrics such as oxygen extraction fraction pushed median discrimination across tumour pairs to about 0.88, compared with about 0.73 for blood volume alone, and crucially that performance held up across acquisition scenarios and in an independent historical validation cohort. Lymphoma versus meningioma was near-perfect; metastasis versus glioblastoma remained the hardest call.

Against that optimism, European Journal of Radiology also publishes a sobering counterweight. Martínez Safar and colleagues systematically reviewed cardiac MRI radiomics and texture analysis for distinguishing the causes of left ventricular hypertrophy — amyloidosis versus hypertrophic cardiomyopathy versus hypertensive heart disease — a differential with genuinely divergent management [10]. Eleven studies, mostly retrospective and single-centre. A validation two-by-two table could be reconstructed in only four of the eleven, no study reported external geographic validation, overfitting or high risk of it was evident in eight of eleven, and certainty was very low throughout. Their conclusion is blunt: the evidence does not yet support clinically credible differentiation, and radiomics here remains adjunctive and hypothesis-generating. Read that alongside the CT radiomics model from Zhao and colleagues in the same journal, which predicts tertiary lymphoid structure status in pancreatic ductal adenocarcinoma across more than 400 patients, with discrimination falling from about 0.89 in training to about 0.72 in the radiogenomic cohort, and with high-scoring patients showing better response and longer progression-free survival on chemoimmunotherapy [8]. The pancreatic work is more rigorous — multiple external and temporal cohorts, a biological correlate in B cell and chemokine gene programmes — but the falling performance across cohorts is exactly the generalisability problem the cardiac review warns about. Treat any single-centre radiomic accuracy figure as a ceiling, not an estimate.

Our third theme is imaging around intervention. In European Journal of Radiology, Linden and colleagues reviewed 560 patients across 862 CT-guided interstitial brachytherapy sessions for liver lesions at a centre that used no routine antibiotic prophylaxis [4]. Liver abscess occurred after about one percent of procedures overall, and in patients without risk factors the rate was well under half a percent. Biliodigestive anastomosis and biliary stents were the clear independent risk factors. Among the small number of patients with enterobiliary manipulation, abscess occurred in 3 of 13 without prophylaxis and 4 of 34 with intensified ciprofloxacin and metronidazole — a numerical trend towards benefit that was not statistically significant, and the numbers are far too small to be definitive. The practical message is reassuring for the majority: routine prophylaxis is hard to justify in patients with an intact biliary system, while those with prior enterobiliary intervention carry a substantially elevated risk that warrants a considered discussion. Then in European Radiology, Vollbrecht and colleagues report a prospective intra-individual cardiac MRI study of 17 cirrhotic patients scanned before and one year after transjugular intrahepatic portosystemic shunt placement [7]. Biventricular ejection fractions stayed preserved and no patient decompensated. Atrial and ventricular volumes and cardiac index rose, left ventricular global longitudinal strain improved, and both T1 and T2 relaxation times fell, indicating resolution of diffuse myocardial oedema — with parallel falls in hepatic and splenic mapping values and liver stiffness. It's a small single-centre cohort, but it reframes post-TIPS cardiac change as adaptive remodelling rather than inevitable deterioration.

Rounding out the week, a prospective randomised protocol study in European Radiology from Zhang and colleagues tested a dual-low thyroid CT strategy in 64 patients [6]. Combining 50-kiloelectron-volt virtual monoenergetic dual-energy imaging with high-strength deep learning reconstruction cut effective radiation dose by about 61 percent, to roughly 0.36 millisieverts, and reduced iodine load by about 20 percent, while delivering lower image noise, steeper edges and higher radiologist quality scores than the standard 120-kilovolt iterative reconstruction protocol. For patients facing lifelong thyroid surveillance, or with borderline renal function, that is a protocol change worth piloting.

If you only have time for one paper this week, make it the deep learning emphysema segmentation study in European Radiology [5]. It directly challenges a quantitative biomarker that sits in routine chest CT reports today, and it does so with multi-centre clinical validation against both radiologist scoring and pulmonary function.

Here are the key takeaways from this week in Radiology. First, pregnancy and lactation are not reasons to defer breast imaging — mammography is safe though less sensitive, ultrasound leads for symptoms, and MRI is an option in lactation but not pregnancy. Second, in high-risk newborns and infants, targeted risk-based ultrasound beats universal screening. Third, threshold-based emphysema quantification is being outperformed by learned segmentation, with better agreement to both radiologists and lung function. Fourth, adding ancillary plaque features to Plaque-RADS, and adding microvascular and metabolic metrics to perfusion imaging, both improve discrimination where conventional single metrics stall. Fifth, be sceptical of radiomics without external validation — the cardiac hypertrophy review found the evidence base very low certainty throughout. And finally, liver abscess after CT-guided brachytherapy is rare unless there has been enterobiliary manipulation, where risk climbs sharply.

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.