This Week in Radiology — Aug 9, 2026

Welcome to This Week in Radiology. This week we're covering 10 notable papers spanning theranostics and molecular imaging, the maturing evidence base for artificial intelligence and structured reporting systems, and a set of protocol-level studies that change how we actually scan patients. Let's dive in.

We start with the biggest headline, and it comes from nuclear medicine. In The Lancet, the PSMAaddition trial reports the second interim analysis of lutetium-177 PSMA-617 moved forward into hormone-sensitive metastatic prostate cancer [1]. This was a phase 3 superiority trial across 169 sites in 20 countries, randomising 1,144 men with PSMA-positive metastatic androgen pathway modulator-naive or sensitive disease to either androgen deprivation plus an androgen receptor pathway inhibitor alone, or that same backbone plus up to six cycles of lutetium PSMA-617 at 7.4 gigabecquerels every six weeks. Every patient had at least one PSMA-positive lesion on centrally read gallium-68 PSMA-11 PET, which is worth emphasising, because central PET reading was the gatekeeper for enrolment. Radiographic progression or death occurred in about a quarter of the radioligand arm versus about 30 percent of the control arm, a 28 percent relative reduction in the risk of progression or death, with median progression-free survival not yet reached in either group. The cost was modest but real: grade 3 or worse adverse events in roughly half the treatment arm versus 43 percent of controls, and dry mouth in 46 percent versus 4 percent. For practising radiologists and nuclear medicine physicians the operational message is immediate. Demand for PSMA PET is going to shift earlier in the disease course, and the quality of that PET read becomes the entry criterion for a therapy that now has phase 3 support in the hormone-sensitive setting. Departments should be planning theranostics capacity and PSMA PET reporting standardisation now, not later.

Our second theme is artificial intelligence in thoracic and breast imaging, where two papers converge on the same conclusion from different angles. In European Radiology, a multicentre case-control study tested two deep learning models for malignancy probability estimation in incidental pulmonary nodules — one trained purely on screening data, one additionally trained on clinical data [3]. Across 269 nodules from three centres, both models achieved areas under the curve of roughly 0.72 to 0.74 versus 0.63 for the Brock model, and at a matched sensitivity of about 78 percent the deep learning models delivered 60 percent specificity compared with 44 percent for Brock. In practical terms, that means meaningfully fewer benign nodules pushed into surveillance or biopsy pathways. Two nuances matter. First, adding clinical training data did not improve performance — the screening-trained model was just as good. Second, this was a cancer-enriched, size-stratified dataset, so the authors are explicit that prospective validation at real-world prevalence is still needed. Alongside this, the European Journal of Radiology published a systematic review and meta-analysis of 42 studies of FDA-cleared and CE-marked artificial intelligence systems for mammography screening [4]. Pooled standalone performance on digital mammography was an area under the curve of 0.89, with sensitivity of about 76 percent and specificity of about 90 percent. The important finding is not the pooled number but where the evidence actually sits: prospective evaluations supported triage and independent or supporting reader configurations, showing non-inferior cancer detection with significant workload reduction against standard double reading. Evidence for fully replacing a radiologist remains lacking, risk of bias was frequently high, and only three of 42 studies were prospective. Read those two papers together and the message is consistent — artificial intelligence is ready to reallocate radiologist attention, not to remove it.

Our third theme is the reality check on structured reporting systems, and here the news is more sobering. In the Journal of Magnetic Resonance Imaging, an individual participant data meta-analysis pooled five studies and 975 observations to measure inter-reader agreement for LI-RADS version 2018 [8]. Agreement for overall categorisation was substantial on MRI, with a kappa of 0.76, but only moderate on CT at 0.50. Agreement on individual major features ranged from poor to good on both modalities, and ancillary features and the LR-M category were worse still, with some kappa values at or below zero. Readers were mostly experienced abdominal subspecialists, so this is close to a best-case scenario. The practical implication is that MRI should be preferred where reproducible LI-RADS categorisation matters, and that ancillary features deserve caution in multidisciplinary discussion. A parallel story comes from the American Journal of Neuroradiology, where two certified neuroradiologists applied the Society of Skeletal Radiology Bone-RADS MRI algorithm to 282 indeterminate vertebral lesions [5]. A Bone-RADS 1 score had a negative predictive value of 100 percent — reassuring, and notable because 80 percent of the original reports on those same lesions had recommended follow-up. Indeterminate categories carried a positive predictive value of only 3.4 percent, falling to zero when the Bone-RADS clinical exclusion criteria were applied. But inter-reader agreement was only moderate at a kappa of 0.60, and the authors argue a vertebra-specific algorithm is needed. So both papers say the same thing: these systems reduce unnecessary follow-up at the benign end, but categorisation is less reproducible than the tables imply. Also from the American Journal of Neuroradiology, arterial spin labeling-derived fractional tumour burden was tested in 86 patients with 102 treated brain metastases to separate recurrence from radiation necrosis [6]. Normalised high fractional tumour burden gave an area under the curve of 0.81 with sensitivity 0.79 and specificity 0.74, and a negative predictive value of 0.88 — significantly better than delta T1, though only comparable to conventional cerebral blood flow metrics rather than superior to them. The appeal is that this is a contrast-free, voxel-level, interpretable biomarker.

Finally, several papers change protocols directly. In the European Journal of Radiology, a study of 446 surgically confirmed parotid lesions asked whether virtual unenhanced dual-energy CT images can replace a true unenhanced acquisition [7]. Virtual unenhanced reproduced lesion attenuation with a bias under 1 Hounsfield unit and met formal equivalence testing overall and across all six histopathological subgroups, with machine learning classifiers performing comparably. The caveats are that background gland attenuation was overestimated by roughly 15 Hounsfield units and lesion detection and border clarity were modestly worse, but omitting the true unenhanced phase would save an estimated 2.4 millisieverts. In the Journal of Magnetic Resonance Imaging, zero echo time MRI was compared prospectively with CT for cervical spine morphometry in 38 surgical patients [9]. Agreement was good to excellent for vertebral body height, disc height and neuroforaminal dimensions, and zero echo time improved depiction of degenerative osseous features over conventional MRI in 92 percent of patients — but pedicle dimensions agreed poorly, with narrow pedicles systematically overestimated, so CT remains necessary when pedicle screw planning is the question. Also in that journal, a prospective non-randomised noninferiority trial of 282 claustrophobic adults found sublingual zolpidem non-inferior to intravenous sedation for MRI completion, at 96.9 percent versus 100 percent, within the pre-specified 10 percent margin, with no respiratory compromise, mild adverse events in about a fifth of patients, and a procedural cost of cents rather than over a hundred dollars [10]. And RadioGraphics offers a practical review on optimising the pelvic MRI protocol for O-RADS MRI, reminding us that the published positive predictive values only apply when the full recommended protocol, including dynamic contrast-enhanced imaging, is actually used [2].

If you only have time for one paper this week, make it the PSMAaddition trial in The Lancet [1]. It moves PSMA radioligand therapy into a far larger patient population and will reshape PSMA PET volumes, reporting standards and theranostics staffing in almost every department.

Here are the key takeaways from this week in Radiology. First, lutetium PSMA-617 added to androgen deprivation and an androgen receptor pathway inhibitor cut radiographic progression or death by 28 percent in hormone-sensitive metastatic prostate cancer, with centrally read PSMA PET as the gatekeeper — expect earlier and higher PSMA PET demand. Second, deep learning beats the Brock model for incidental pulmonary nodule risk, and commercial mammography artificial intelligence has its strongest prospective evidence in triage and supporting-reader roles, not in replacing radiologists. Third, LI-RADS categorisation is more reproducible on MRI than CT, and Bone-RADS in the spine has only moderate inter-reader agreement — treat the benign categories as reliable and the indeterminate ones with caution. Fourth, virtual unenhanced dual-energy CT can safely replace the true unenhanced phase in selected parotid imaging, saving about 2.4 millisieverts, and zero echo time MRI can substitute for CT in cervical morphometry except for pedicle measurement. And fifth, sublingual zolpidem is a low-cost, non-inferior alternative to intravenous sedation for selected claustrophobic patients.

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.