This week in Radiology

We start with a large negative trial that matters for anyone involved in post-thrombectomy care. In the BMJ, Yuan and colleagues report HeadSOAR, a multicentre randomised trial across 67 comprehensive stroke centres in China enrolling 1,368 adults with anterior circulation large vessel occlusion who achieved successful reperfusion after thrombectomy. Patients were randomised to keep the head elevated at 30 to 40 degrees or flat at 0 to 10 degrees for the first 72 hours, and the primary outcome was the full distribution of modified Rankin scores at 90 days. The median Rankin score was 3 in both arms, and the shift analysis showed no significant benefit from head elevation. Mortality at 90 days was about 18 percent with elevation versus about 20 percent flat, again not a significant difference. The authors are careful to note the between-group difference was smaller than they had powered for, so a modest benefit can't be excluded, but as it stands there is no evidence that mandating head elevation after successful reperfusion improves outcomes. For practice, this removes an argument for rigid positioning protocols and lets positioning be driven by aspiration risk, comfort and haemodynamics.

The largest thread this week is artificial intelligence and quantitative imaging as a diagnostic safety net, and the headline paper comes from Nature Medicine. Zhang and colleagues built LiON, a contrast-enhanced CT system for liver malignancy diagnosis, trained on roughly 6,400 patients and validated retrospectively across more than 22,000 across multicentre and real-world cohorts. Discrimination for malignancy was high, with an area under the curve of about 0.98 overall, holding up in patients with hepatic steatosis and dropping only modestly, to about 0.92, in cirrhosis, which is where most of us find these cases hardest. They then ran a single-arm prospective trial in more than 10,000 patients with the model acting as an additional reader inside the existing workflow. It met its endpoint with an area under the curve of about 0.95, and importantly it flagged 51 lesions that had been overlooked, 15 of them malignant, leading to 37 amended reports and 22 multidisciplinary escalations. That is the most concrete demonstration to date of an artificial intelligence second reader changing real reports, but note the design: single-arm, no comparator, so we cannot say how many of those catches a second human reader would have made.

If you only have time for one paper this week, make it the LiON study in Nature Medicine [2]. It is the largest prospective deployment of an artificial intelligence reader inside a live radiology workflow to date, and it documents amended reports and changed management rather than just an area under the curve.

Two papers round out the week on procedural safety and reporting behaviour. Xu and colleagues, again in European Radiology, pooled a decade of experience across seven centres on percutaneous pulsed electric field ablation for hepatocellular carcinoma in high-risk anatomical locations. In 304 patients, the overall complication rate was under 8 percent and major complications occurred in about 3 percent, mostly perioperative systemic events, thoracic injury with subphrenic tumours, and needle tract complications. Post-ablation syndrome affected roughly one in five patients, lower platelet count predicted major complications, and after propensity matching microsecond and nanosecond techniques had comparable safety. Meanwhile in the Journal of the American College of Radiology, Marsiglia and colleagues looked at recommendations for additional orbital imaging in a health system's head and neck CT and MRI reports over a year. Such recommendations were rare, appearing in about two in a thousand reports, and only about a third of those patients had dedicated orbital imaging within a year. But among those who did, 93 percent had pathology, most commonly neoplastic. The authors caution that some of that imaging may have been planned surveillance rather than a response to the recommendation, but the yield is high enough to argue for closing the loop on these recommendations more reliably.

Quantitative imaging biomarkers dominate the rest of this theme, and three of them appear in European Radiology. Shen and colleagues prospectively compared ultrasound-derived fat fraction against liver biopsy in 418 patients with suspected or confirmed metabolic dysfunction-associated steatotic liver disease across two centres. Correlation with histology was strong, bias against histopathology was under two percent, and they propose practical integer cut-offs of 6 percent, 15 percent and 21 percent for mild, moderate and severe steatosis, with discrimination for mild steatosis around 0.97 and for severe steatosis above 0.9 on validation. Their dual-threshold strategy, with a deliberate indeterminate zone, cut ambiguous classifications by up to about a quarter. The caveat is a biopsy-selected cohort, few severe cases, and no inter-observer data, so treat those thresholds as promising rather than definitive. In pancreatic cancer, Holland and colleagues prospectively studied quantitative dynamic contrast-enhanced MRI in 42 patients scanned at baseline and again about seven weeks into chemotherapy, with a portable perfusion phantom scanned alongside to correct scanner-dependent error. Early change in Ktrans separated responders from non-responders with roughly 80 percent accuracy uncorrected, rising to about 98 percent with phantom correction using the extended Tofts model, at a point when conventional size-based CT assessment identified response in fewer than one in five patients. Those are small numbers from two institutions and the near-perfect accuracy will need external replication, but the principle, that perfusion change precedes shrinkage in pancreatic ductal adenocarcinoma, is clinically actionable. And in lung cancer, Zong and colleagues developed RaFiST, a contrast-enhanced CT radiomics model trained in 532 patients with resected non-small cell lung cancer to predict histologically quantified tumour fibrosis. Fibrosis score was an independent predictor of both disease-free and overall survival, the model predicted high fibrosis with an area under the curve around 0.88 in training and 0.81 externally, and combining it with clinical variables gave the best five-year disease-free survival prediction. Transcriptomics linked high fibrosis to extracellular matrix remodelling, hypoxia and immune suppression, which gives the imaging signature some biological grounding.

Quantitative imaging biomarkers dominate the rest of this theme, and three of them appear in European Radiology. Shen and colleagues prospectively compared ultrasound-derived fat fraction against liver biopsy in 418 patients with suspected or confirmed metabolic dysfunction-associated steatotic liver disease across two centres. Correlation with histology was strong, bias against histopathology was under two percent, and they propose practical integer cut-offs of 6 percent, 15 percent and 21 percent for mild, moderate and severe steatosis, with discrimination for mild steatosis around 0.97 and for severe steatosis above 0.9 on validation. Their dual-threshold strategy, with a deliberate indeterminate zone, cut ambiguous classifications by up to about a quarter. The caveat is a biopsy-selected cohort, few severe cases, and no inter-observer data, so treat those thresholds as promising rather than definitive. In pancreatic cancer, Holland and colleagues prospectively studied quantitative dynamic contrast-enhanced MRI in 42 patients scanned at baseline and again about seven weeks into chemotherapy, with a portable perfusion phantom scanned alongside to correct scanner-dependent error. Early change in Ktrans separated responders from non-responders with roughly 80 percent accuracy uncorrected, rising to about 98 percent with phantom correction using the extended Tofts model, at a point when conventional size-based CT assessment identified response in fewer than one in five patients. Those are small numbers from two institutions and the near-perfect accuracy will need external replication, but the principle, that perfusion change precedes shrinkage in pancreatic ductal adenocarcinoma, is clinically actionable. And in lung cancer, Zong and colleagues developed RaFiST, a contrast-enhanced CT radiomics model trained in 532 patients with resected non-small cell lung cancer to predict histologically quantified tumour fibrosis. Fibrosis score was an independent predictor of both disease-free and overall survival, the model predicted high fibrosis with an area under the curve around 0.88 in training and 0.81 externally, and combining it with clinical variables gave the best five-year disease-free survival prediction. Transcriptomics linked high fibrosis to extracellular matrix remodelling, hypoxia and immune suppression, which gives the imaging signature some biological grounding.

A third theme is surveillance and hardware. Also in European Radiology, Yang and colleagues addressed a problem created by the Liver Imaging Reporting and Data System 2024 surveillance algorithm, which now offers abbreviated MRI for cirrhotic patients whose ultrasound visualisation is severely limited. In 264 Child-Pugh A or B cirrhotic patients with a prior ultrasound visualisation score of C, noncontrast abbreviated MRI gave adequate visualisation in over 91 percent of cases on diffusion-weighted imaging and over 95 percent on T2-weighted imaging. Eighteen hepatocellular carcinomas were detected and 17 of those were early stage. It's retrospective and single-institution, but it supports switching these patients out of ultrasound surveillance rather than repeating inadequate studies. On the hardware side, Zou and colleagues report the first performance test of a dual-source CT system combining two 16-centimetre detectors with two tubes on one gantry, along with a purpose-built cross-scatter correction. Spatial resolution and low-contrast detectability were preserved after correction, phantom testing showed motion-free imaging at 100 beats per minute and under simulated arrhythmia, and a patient scanned at a mean heart rate of 197 beats per minute yielded sharp coronary arteries at 1.48 millisieverts. This is a phantom-and-illustration study, not a clinical trial, but it addresses a real trade-off between z-axis coverage and temporal resolution.

A third theme is surveillance and hardware. Also in European Radiology, Yang and colleagues addressed a problem created by the Liver Imaging Reporting and Data System 2024 surveillance algorithm, which now offers abbreviated MRI for cirrhotic patients whose ultrasound visualisation is severely limited. In 264 Child-Pugh A or B cirrhotic patients with a prior ultrasound visualisation score of C, noncontrast abbreviated MRI gave adequate visualisation in over 91 percent of cases on diffusion-weighted imaging and over 95 percent on T2-weighted imaging. Eighteen hepatocellular carcinomas were detected and 17 of those were early stage. It's retrospective and single-institution, but it supports switching these patients out of ultrasound surveillance rather than repeating inadequate studies. On the hardware side, Zou and colleagues report the first performance test of a dual-source CT system combining two 16-centimetre detectors with two tubes on one gantry, along with a purpose-built cross-scatter correction. Spatial resolution and low-contrast detectability were preserved after correction, phantom testing showed motion-free imaging at 100 beats per minute and under simulated arrhythmia, and a patient scanned at a mean heart rate of 197 beats per minute yielded sharp coronary arteries at 1.48 millisieverts. This is a phantom-and-illustration study, not a clinical trial, but it addresses a real trade-off between z-axis coverage and temporal resolution.

Two papers round out the week on procedural safety and reporting behaviour. Xu and colleagues, again in European Radiology, pooled a decade of experience across seven centres on percutaneous pulsed electric field ablation for hepatocellular carcinoma in high-risk anatomical locations. In 304 patients, the overall complication rate was under 8 percent and major complications occurred in about 3 percent, mostly perioperative systemic events, thoracic injury with subphrenic tumours, and needle tract complications. Post-ablation syndrome affected roughly one in five patients, lower platelet count predicted major complications, and after propensity matching microsecond and nanosecond techniques had comparable safety. Meanwhile in the Journal of the American College of Radiology, Marsiglia and colleagues looked at recommendations for additional orbital imaging in a health system's head and neck CT and MRI reports over a year. Such recommendations were rare, appearing in about two in a thousand reports, and only about a third of those patients had dedicated orbital imaging within a year. But among those who did, 93 percent had pathology, most commonly neoplastic. The authors caution that some of that imaging may have been planned surveillance rather than a response to the recommendation, but the yield is high enough to argue for closing the loop on these recommendations more reliably.

Quantitative imaging biomarkers dominate the rest of this theme, and three of them appear in European Radiology. Shen and colleagues prospectively compared ultrasound-derived fat fraction against liver biopsy in 418 patients with suspected or confirmed metabolic dysfunction-associated steatotic liver disease across two centres. Correlation with histology was strong, bias against histopathology was under two percent, and they propose practical integer cut-offs of 6 percent, 15 percent and 21 percent for mild, moderate and severe steatosis, with discrimination for mild steatosis around 0.97 and for severe steatosis above 0.9 on validation. Their dual-threshold strategy, with a deliberate indeterminate zone, cut ambiguous classifications by up to about a quarter. The caveat is a biopsy-selected cohort, few severe cases, and no inter-observer data, so treat those thresholds as promising rather than definitive. In pancreatic cancer, Holland and colleagues prospectively studied quantitative dynamic contrast-enhanced MRI in 42 patients scanned at baseline and again about seven weeks into chemotherapy, with a portable perfusion phantom scanned alongside to correct scanner-dependent error. Early change in Ktrans separated responders from non-responders with roughly 80 percent accuracy uncorrected, rising to about 98 percent with phantom correction using the extended Tofts model, at a point when conventional size-based CT assessment identified response in fewer than one in five patients. Those are small numbers from two institutions and the near-perfect accuracy will need external replication, but the principle, that perfusion change precedes shrinkage in pancreatic ductal adenocarcinoma, is clinically actionable. And in lung cancer, Zong and colleagues developed RaFiST, a contrast-enhanced CT radiomics model trained in 532 patients with resected non-small cell lung cancer to predict histologically quantified tumour fibrosis. Fibrosis score was an independent predictor of both disease-free and overall survival, the model predicted high fibrosis with an area under the curve around 0.88 in training and 0.81 externally, and combining it with clinical variables gave the best five-year disease-free survival prediction. Transcriptomics linked high fibrosis to extracellular matrix remodelling, hypoxia and immune suppression, which gives the imaging signature some biological grounding.