This Week in Radiology — Jun 23, 2026

Welcome to This Week in Radiology. This week we're covering 9 notable papers spanning neuroimaging advances in stroke and fetal medicine, screening and artificial intelligence validation in oncology, and protocol optimizations in photon-counting CT and MR cholangiopancreatography. Let's dive in.

We start in neuroimaging, where we see important developments in stroke prognostication, fetal brain monitoring, and tumor phenotyping. In patients presenting with acute ischemic stroke due to anterior circulation large vessel occlusion, predicting long-term outcomes remains a complex challenge. A study published in the American Journal of Neuroradiology investigated the Perfusion Collateral Impairment Score, or PCIS, evaluating its prognostic value over standard imaging parameters in two hundred and eighty-three patients [3]. The researchers found that adding the ordinal PCIS to standard imaging models—such as those utilizing Tan, COVES, or ASPECTS scores—was independently associated with ninety-day outcomes, with roughly half of the cohort experiencing an unfavorable outcome. Interestingly, when adjusting for age and baseline stroke severity, only the model combining COVES and PCIS retained a statistically significant independent association. While the standalone discriminative ability of the PCIS was modest, it appears to capture unique hemodynamic properties that complement traditional CT angiography collateral scores, suggesting it could serve as a valuable adjunctive tool in acute stroke workups.

Moving from acute stroke to spine imaging, we encounter another diagnostic challenge: identifying cerebrospinal fluid-venous fistulas, which are a common cause of spontaneous intracranial hypotension. These fistulas are notoriously difficult to detect on CT myelography because they do not present with pooling contrast, instead relying on the visualization of minute amounts of contrast within tiny paraspinal veins. To address the diagnostic uncertainty that often accompanies these cases, a team writing in the American Journal of Neuroradiology demonstrated the real-world application of the Duke CSF-Venous Fistula Confidence Score [2]. This structured reporting system is anchored in specific imaging findings to help radiologists communicate their level of certainty regarding the presence or absence of a fistula at individual spinal levels, providing a much-needed framework to standardize myelogram interpretations.

In fetal neuroimaging, a retrospective cohort study published in European Radiology evaluated the role of serial prenatal MRI in monitoring congenital cytomegalovirus infection [5]. The investigators compared seventy-seven fetuses, fifteen of whose mothers were treated with oral valacyclovir and sixty-two who were untreated. They focused on the development of minor brain lesions, such as temporal pole T2 hyperintensity and small cysts. The results were striking: nearly three-quarters of the treated fetuses maintained persistently normal prenatal MRI findings, compared to just over thirty-seven percent of the untreated group, representing an almost five-fold increase in the likelihood of a normal brain MRI. Early-onset minor lesions were observed exclusively in the untreated group. These findings suggest that maternal valacyclovir therapy exerts a protective effect on the fetal brain and highlights the potential of serial fetal MRI to serve as an imaging surrogate marker of treatment response.

Rounding out our neuroimaging coverage, a study in the Journal of Magnetic Resonance Imaging proposes an expansion of a familiar brain tumor biomarker [7]. The classic T2-FLAIR mismatch sign is highly specific for isocitrate dehydrogenase-mutant astrocytomas, but its strict definition limits how often it can be applied. The researchers evaluated an expanded T2-FLAIR mismatch phenotype, which incorporates spatially heterogeneous signals, in a cohort of three hundred and forty-nine patients. This expanded phenotype was identified in about one-third of the patients, more than doubling the detection rate of the classic sign, which was seen in only about fourteen percent of cases. Patients with the expanded phenotype demonstrated significantly more favorable overall survival, with an absolute survival difference of about three months over a thirty-six-month period. Although the expanded phenotype was not an independent prognostic factor in multivariable analysis, it successfully delineates a subgroup of patients with favorable clinicopathologic features and broadens the clinical utility of this important imaging biomarker.

Next, we turn to oncology and the evolving role of screening and artificial intelligence. In the realm of prostate cancer screening, the debate over the utility of prostate-specific antigen, or PSA, testing continues. A secondary analysis of the landmark Stockholm3-MRI randomized screening trial, published in the Annals of Internal Medicine, evaluated a multivariable risk score that combines PSA, plasma protein biomarkers, polygenic risk, and clinical factors [1]. Among over twelve thousand Swedish men, the Stockholm3 score provided a substantially higher clinical net benefit than PSA alone for detecting clinically significant prostate cancer over a two-year follow-up. Stockholm3 achieved a sensitivity of ninety percent compared to seventy-four percent for PSA, while maintaining a similar specificity of around ninety percent. Crucially, the Stockholm3 approach cut the false-negative rate to ten percent, compared to twenty-six percent for PSA, meaning it missed far fewer clinically significant cancers while avoiding unnecessary biopsies. Despite limitations like a twenty-five percent participation rate and a predominantly European cohort, these findings make a strong case for replacing standalone PSA testing with multivariable risk models.

In breast oncology, predicting axillary lymph node metastasis is vital for staging and treatment planning, yet conventional imaging remains highly operator-dependent. A systematic review and meta-analysis published in European Radiology pooled data from forty-one studies to evaluate how well deep learning and hand-crafted radiomics models perform using MRI and ultrasound [9]. Across these studies, artificial intelligence models demonstrated moderate diagnostic accuracy, with an area under the curve of zero point eight four on internal validation and zero point eight two on external validation. The authors noted that standalone clinical utility remains limited, as indicated by a positive likelihood ratio of three and a negative likelihood ratio of zero point three three. However, when deep learning and hand-crafted radiomics were combined into ensemble models, performance improved significantly, reaching an area under the curve of zero point nine two on ultrasound and zero point eight eight on MRI. These findings suggest that while these tools are not ready to replace invasive biopsies, they can play a powerful adjunctive role in risk stratification.

However, as we implement these artificial intelligence and radiomics models, we must remain highly critical of how they are validated. A cautionary study also published in European Radiology highlights a major methodological pitfall in the field [6]. Researchers trained three conditional deep generative models on fifty public radiomic datasets to see if standard synthetic data quality metrics could predict whether data augmentation actually improves model performance under simulated external validation. They discovered a classic ecological fallacy: while aggregate quality metrics correlated significantly with performance gains across the entire dataset pool, these correlations completely vanished when looking within individual generators. In fact, selecting generators based on these quality metrics actually underperformed models trained on real data alone. This critical finding warns us that current synthetic data quality metrics are poor proxies for actual clinical utility, and task-specific external validation remains absolutely indispensable before any model is deployed in a clinical setting.

Our final theme this week focuses on advanced imaging protocols and reconstruction algorithms. In temporal bone imaging, achieving high spatial resolution without exposing patients to excessive radiation is a constant balancing act. A prospective study in the European Journal of Radiology evaluated the clinical impact of low-dose, ultra-high-resolution temporal bone imaging using photon-counting detector CT with one hundred kilovolt-peak tin filtration [4]. Comparing forty-six patients who underwent this protocol against their prior conventional energy-integrating detector CT scans, the researchers reported a fifty-three percent reduction in radiation dose. Despite the lower dose, the photon-counting protocol reconstructed at zero point six millimeters demonstrated significantly lower image noise and higher contrast-to-noise ratios. Furthermore, ultra-thin zero point two millimeter reconstructions, despite having higher noise, leveraged the superior spatial resolution of photon-counting technology to provide vastly superior visualization of fine anatomical structures, marking a major step forward for low-dose skull base imaging.

Lastly, we look at pancreatic imaging, where the debate between two-dimensional and three-dimensional magnetic resonance cholangiopancreatography, or MRCP, remains unresolved. A prospective study of one hundred and seventy-four patients in the European Journal of Radiology compared three-dimensional MRCP using super-resolution deep learning reconstruction against conventional two-dimensional MRCP with compressed sensing and two-dimensional MRCP with the same super-resolution reconstruction [8]. While the three-dimensional technique yielded the highest signal-to-noise and contrast-to-noise ratios for the distal common bile duct, both two-dimensional techniques outperformed it in overall image quality, motion artifact reduction, and main pancreatic duct visualization. Notably, for the detection of focal pancreatic duct narrowing without upstream dilation, the three-dimensional technique suffered from a high false-positive rate of over forty percent, whereas neither two-dimensional technique produced any false-positive findings. This suggests that two-dimensional MRCP enhanced with super-resolution deep learning reconstruction may offer the most reliable visualization of the main pancreatic duct in clinical practice.

If you only have time for one paper this week, make it the Stockholm3-MRI screening study published in the Annals of Internal Medicine [1]. This paper provides compelling, high-impact evidence that replacing traditional PSA screening with a multivariable risk score combined with MRI can dramatically reduce false negatives and unnecessary biopsies, offering a clear blueprint for the future of population-based prostate cancer screening.

Here are the key takeaways from this week in Radiology: First, the Stockholm3 multivariable risk score, when paired with MRI, significantly outperforms PSA alone for prostate cancer screening, reducing false-negative rates from twenty-six percent to ten percent. Second, serial prenatal MRI can serve as an effective surrogate marker of treatment response in congenital cytomegalovirus, demonstrating a clear association between maternal valacyclovir therapy and normal fetal brain development. Third, caution is required when validating artificial intelligence models; synthetic radiomics quality metrics suffer from an ecological fallacy and do not predict model performance under external validation, emphasizing that direct external testing remains mandatory. Fourth, photon-counting detector CT with tin filtration can cut radiation doses for temporal bone imaging by more than half while simultaneously improving spatial resolution and image quality. And finally, super-resolution deep learning-enhanced two-dimensional MRCP provides superior main pancreatic duct visualization and fewer false positives for focal narrowing compared to three-dimensional techniques.

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 for audio briefings on your own clinical questions and papers, visit audioscholar dot C C.

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