This Week in Radiology — May 21, 2026

Welcome to This Week in Radiology. This week we're covering 10 notable papers spanning the rapid evolution of artificial intelligence in our field, key updates in neurointerventional and spine procedures, and new frontiers in diagnostic imaging and safety. Let's dive in. First, we turn to artificial intelligence, where four papers this week tackle everything from standardizing research to long-term risk prediction and the practicalities of using large language models in the reading room. To address the heterogeneity of AI development, the AJNR. American journal of neuroradiology has published new study-specific guidelines. These aim to standardize reporting for different types of AI studies, such as those for classification, segmentation, or workflow optimization. The goal is to enforce rigor, particularly around reference standards and clinical relevance, to help bridge the gap between algorithms and actual practice [2]. Moving from guidelines to application, a study in Science translational medicine presents a new AI model for long-term breast cancer risk prediction [7]. Developed in Sweden and validated in two large United States cohorts, this model uses mammographic images to estimate 10-year absolute risk. Results In validation cohorts, the model achieved a 10-year AUC of 0.72. More impressively, for individuals in the top 10% of risk, this new AI tool predicted 33% of incident breast cancers. This significantly outperformed existing clinical tools like Tyrer-Cuzick, which predicted 23%, and another image-based AI tool, Mirai, which predicted 24%. But how do we interact with these models? A study in AJR. American journal of roentgenology explored a human-in-the-loop workflow using a large language model, or LLM, for thoracic imaging diagnostics [10]. Methods Radiologists first interpreted cases and wrote free-text descriptions of their findings. This text, without the images, was fed to an LLM, which then ranked a list of differential diagnoses. The radiologists then reviewed the LLM's output and could revise their diagnosis. Results Interestingly, the LLM was more accurate when using the radiologists' text descriptions than when analyzing the images directly. The workflow improved diagnostic accuracy for both attending thoracic radiologists and residents. However, the study uncovered a critical nuance: residents saw a larger accuracy boost but were also much more likely to be swayed by a misleading LLM suggestion, switching to an incorrect diagnosis over 60% of the time, compared to attendings who did so only about 32% of the time. This highlights how expertise shapes our interaction with AI, influencing both how we prompt the model and how we critically evaluate its output. Finally, looking at the basic science frontier, a paper in Nature introduces MouseMapper, a deep-learning framework for whole-body analysis in mice at a cellular level [5]. This tool can automatically segment 31 organs and quantify things like nerve fibers and immune cell clusters across the entire animal. Using it to study diet-induced obesity, the researchers identified specific structural damage to facial nerves and mapped out inflammation across different tissues. This provides a powerful, scalable approach for linking molecular changes in animal models to systemic diseases. Next, we have three papers from AJNR. American journal of neuroradiology focusing on the efficacy and safety of common neurointerventional and spine procedures. For treating wide-neck intracranial aneurysms, stent-assisted coiling is a common approach. A multicenter study provides the first long-term data on the LVIS EVO stent [1]. The Study Investigators followed 158 patients for at least 12 months. Results At final follow-up, 85.4% of aneurysms showed complete occlusion, and 96.8% had adequate occlusion. Recanalization occurred in about 8% of patients, with only 1.3% requiring retreatment. The procedure carried a morbidity rate of 4.4%. These findings support the long-term safety and durability of this device for complex aneurysms. Shifting to procedural safety, another AJNR paper details a quality improvement project to prevent wrong-site vertebroplasty, a rare but devastating 'never' event [3]. Methods The team implemented several iterative changes. The primary goal was to ensure a fluoroscopy image marking the correct vertebral level with forceps was obtained and saved before the procedure. Conclusions After several cycles of adjustment, they achieved near-perfect compliance. A key intervention was the introduction of a 'second time-out,' where a technologist would remind the physician to obtain and save the localization image. Importantly, these new safety steps did not significantly increase procedure time, fluoroscopy time, or sedation time. This provides a simple, effective template for improving safety in image-guided spine procedures. Our final paper in this section, also from AJNR, offers a fascinating observation on epidural injections [6]. The study looked at patients undergoing multilevel epidural blood patching. The Study After injecting iodinated contrast into the epidural space, the investigators performed CT scans that included the kidneys. Results They found that 100% of patients showed contrast in their renal collecting systems, with an average time of just 23 minutes from the first injection. Conclusions This provides clear imaging evidence of rapid systemic absorption from the epidural space. The key clinical takeaway is that seeing contrast in the kidneys after a myelogram or other epidural injection should not, on its own, be mistaken for an indirect sign of a CSF leak or other pathology. It is a normal physiologic finding. Our final section covers important findings in breast imaging, contrast safety in a vulnerable population, and a novel therapeutic use of ultrasound. In European radiology, a large prospective study evaluated multiparametric breast MRI for non-mass enhancement, or NME, a common diagnostic challenge [4]. Methods The protocol combined standard dynamic contrast-enhanced imaging with diffusion-weighted imaging, or DWI, and ADC measurements. Results Across 351 women, this combined approach demonstrated excellent performance, with 98.4% sensitivity and 86.4% specificity. The most important metric may be the negative predictive value, which was nearly 98%. This high NPV suggests that multiparametric MRI can be used confidently to rule out malignancy in NME lesions, potentially helping to avoid a significant number of unnecessary biopsies. Next, a study in AJR. American journal of roentgenology tackles the important question of contrast-induced acute kidney injury, or AKI, in neonates in the ICU [9]. The Study Using robust statistical methods including propensity-score matching to minimize confounding, researchers compared neonates who received iodinated contrast for a CT scan to a control group who did not. AKI was defined using modern, neonatal-specific criteria. Results The results showed that contrast exposure was associated with a higher risk of AKI. In the matched cohort, the incidence was 12.3% in the contrast group versus 7.0% in the control group, representing a risk increase of nearly 90%. However, this increase was confined to stage 1, or mild, AKI. There was no significant difference in the rates of more severe stage 2 or 3 AKI. This provides crucial, controlled evidence that refines our understanding of contrast risk in this fragile population. Finally, from Science translational medicine, a study explores transcranial ultrasound stimulation, or TUS, as a noninvasive therapy for Parkinson's disease [8]. Methods Researchers used MRI-guided TUS to target specific brain regions in patients with Parkinson's who already had deep brain stimulation electrodes implanted in the subthalamic nucleus, or STN. This allowed for direct measurement of the neural effects of the ultrasound. Results The effects were highly target-specific. When TUS was aimed at the primary motor cortex, it reduced pathological beta wave activity in the STN and was associated with improved motor signs. In contrast, stimulating the globus pallidus internus actually increased beta activity and did not improve motor function. This study provides direct mechanistic evidence that TUS can safely and selectively modulate deep brain circuits, supporting its potential as a future noninvasive therapeutic tool. If you only have time for one paper this week, make it the study in AJR. American journal of roentgenology on using large language models for diagnostic assistance in thoracic imaging [10]. It's a must-read because it moves beyond AI performance metrics to explore the complex, dynamic interaction between the human radiologist and the model, showing how our own expertise is critical in both using and evaluating these emerging tools. Here are the key takeaways from this week in Radiology. First, in breast MRI, adding diffusion-weighted imaging to the standard protocol for non-mass enhancement offers a very high negative predictive value, giving you more confidence to avoid biopsy for lesions with benign features. Second, for neonates in the ICU, iodinated contrast is associated with an increased risk of mild, stage 1 acute kidney injury, but not more severe forms. This helps in risk-benefit discussions for necessary imaging in this group. Third, when using large language models to help with differential diagnoses, be mindful that expertise matters. Trainees may benefit more but are also more susceptible to being misled, underscoring the need for critical evaluation of AI output. And finally, a couple of quick pearls for spine procedures: implementing a simple 'second time-out' can be highly effective for preventing wrong-site errors. And remember that contrast injected into the epidural space is absorbed systemically within minutes, so seeing it in the kidneys is a normal finding, not a sign of a leak. 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.

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