Weekly AI digest
Radiology & medical imaging AI · week of 18 to 24 October 2026
165 peer-reviewed papers · 6 industry and regulatory items · conference and KOL highlights. Adapted from my weekly intelligence report. Full report: FR (PDF) · EN (PDF) · RU (PDF).
Governance takeaway
What to do about it this week
Measure the pair and the place, not the model. This week a randomised crossover experiment in The Lancet Digital Health showed that ignoring correct AI advice damaged physicians' diagnostic performance more than following incorrect advice, with explanations helping mostly on hard positive cases; a prospective community screening study showed a high-AUROC triage tool delivering a 17.6% positive predictive value because disease was rare in the field; and a Turkish reporting-LLM study showed perfectly parsed output hiding incomplete clinical content.
Add two lines to your acceptance-test template: a reliance measure (how readers respond to deliberately mixed correct and incorrect AI output) and a prevalence-matched PPV with expected alert volume. Keep both as monthly monitoring KPIs after go-live.
Regulation & AI Act watch
Elucid FFR-IQ cleared for FFR-CT from coronary CT angiography
FDA 510(k) for noninvasive fractional flow reserve estimation that folds histology-validated plaque quantification into the physiology estimate; the 92.2% sensitivity and 93.1% specificity come from the vendor's own ELUCIDATE study, cited to the 510(k) summary rather than a peer-reviewed paper. First question for the vendor: how does the tool handle the 0.75 to 0.80 gray zone, and what is its rejection rate on your CCTA quality.3DR Labs NeuroDR cleared for autonomous neurovascular 3D post-processing
A 510(k) for automated segmentation, modeling and centerlining of head and neck CTA, returned to PACS with no technologist clicks. This is a visualization clearance, not a detection or triage claim, and no accuracy study is cited. First question: how are failed or implausible reconstructions flagged before anyone treats them as diagnostic-ready.RIVANNA BoneEnhance cleared as an ML layer on Accuro XV musculoskeletal ultrasound
Bone and soft-tissue segmentation with 3D visualization during review, validated on five independent test sites (233 volumes, 69,900 images, 173 subjects) separate from three training sites; anatomy labeling only, the injury-detection module is investigational. First question: performance by body habitus, age and postsurgical anatomy.Coreline Soft AVIEW Aorta Metrics cleared for aortic quantification on chest CT
Automated aortic segmentation and zone-wise diameter measurement, a quantification tool rather than an aneurysm detection claim, announced without performance figures. First question: agreement with expert manual measurement on your scanners, and whether its diameters match the conventions your surveillance thresholds assume.NeuroXT NEXT-Cue cleared to support prognosis in mild cognitive impairment
Class II 510(k) for software combining brain MRI, demographics and cognitive scores to inform assessment of five-year progression to Alzheimer's dementia; the announcement cites no performance data. First question: do the validation cohorts look like a real memory clinic, and how is calibration reported by age, sex and education.Optellum lung nodule AI joins the RapidAI enterprise platform
A distribution and integration deal that changes the deployment route, not the evidence: clearance and Medicare reimbursement are unchanged. First question: who owns and can export the monitoring data the platform collects, and how are model versions controlled when two vendors update independently.
Evidence you can use
Selected from 165 papers indexed in PubMed this week.
Development of a Novel Real-Time Computer-Assisted Colonoscopy Diagnostic Tool for Colorectal Neoplasia and Cancer to Enable Personalized Patient Management: Protocol for a Multicenter Proof-of-Concept Study
JMIR Research Protocols, 21 October 2026Mobile AI-Assisted Oral Lesion Triage for Community Oral Cancer Screening: Prospective Field Evaluation Study
JMIR mHealth and uHealth, 20 October 2026TurkRadLLM: A Compact MedGemma-Based Model for Structuring Turkish Lumbar and Cervical Spine MRI Reports-A Multi-reader Validation Study
Journal of Imaging Informatics in Medicine, 20 October 2026Reducing Annotation Costs Using Knowledge Distillation for Label‑Efficient Negation Detection in Radiology Reports: Multisite Internal Validation Study
JMIR Medical Informatics, 22 October 2026An interpretable contrast-enhanced CT-based cross-attention transformer model for predicting progression-free survival in muscle-invasive bladder cancer: a multicenter study
Abdominal Radiology, 22 October 2026Automated carotid artery atherosclerotic inflammation segmentation in PET using MSR-Net
Physics in Medicine and Biology, 21 October 2026Frozen Echocardiography Foundation Model Embeddings Contain Linearly Recoverable LVOT VTI and TAPSE Report-Label Signal in MIMIC-IV-ECHO
Journal of Imaging Informatics in Medicine, 19 October 2026Examining reliance patterns on AI advice in medical imaging: a mixed-methods, randomised, crossover experiment
The Lancet Digital Health, 23 October 2026
Post-market signal
An Emory-led study in npj Digital Medicine measured two FDA-cleared pulmonary embolism algorithms across 67,869 CT scans in a 17-facility health system. The CTPA triage model reached 86.8% sensitivity and the incidental-PE model 73.5%, both below their clearance benchmarks, while specificity exceeded cleared levels. Sensitivity fell to 72.9% for subsegmental and 65.3% for non-acute emboli. Labels came from reports written with the AI output visible, so the true gap may be wider. A negative flag cannot end the search for small or chronic clots.
The same pattern appeared in RSNA session material reported this week by Diagnostic Imaging Europe: real-world intracranial hemorrhage triage sensitivity of about 82% against 96% in the clearance data presented, and a 227-radiologist study in which the average effect of AI assistance was zero while individual readers gained or lost heavily. A Lancet Medical Imaging and Theranostics commentary adds the structural reason: scanner generations, disease definitions and therapies keep moving the data away from what the model was validated on, so monitoring and revalidation have to be designed in.
Playbook snippet
One step for your AI committee
Anchor: EU AI Act Article 10 (data governance).
Step: Add one clause to your next imaging-AI RFP: the supplier discloses the demographic and scanner composition of training and test data, and reports performance by subgroup. Record non-disclosure as a risk item under Article 9 rather than treating silence as acceptable.
Why: Diversity, not volume, is the specification worth writing in: a published bias audit found the model trained on the most homogeneous cohort was both the worst and the most biased, while a moderately sized but diverse cohort beat a much larger one (Lee RS et al., AJNR, PMID 41663205, DOI 10.3174/ajnr.A9217).
Worked example: This week's multicenter colonoscopy CADx protocol (Olabintan O et al., JMIR Research Protocols, PMID 42905971, DOI 10.2196/77277) designs composition in from the start: more than 17,000 images from eight European centers and one Japanese center, and subgroup performance predefined as a secondary outcome. Demand that level of composition reporting from vendors in your own procurement and audits.
From the field
Regulators and societies spent the week on generative AI. The FDA comment period on its generative AI discussion paper closed on 20 October with more than 500 submissions, and the FDA's digital health director said draft guidance will follow within months. RSNA and ACR both argued that one-time premarket review cannot cover models that change, called for continuous real-world monitoring, and asked that failures be measured as clinically consequential errors such as false assertions and omissions rather than generic hallucination rates. Hospital and physician groups pressed for developers, not providers, to carry the monitoring burden.
In Europe, the Royal College of Radiologists published the second edition of its guidance on integrating AI into RIS and PACS reporting workflows, and the European Society of Emergency Radiology's teleradiology committee flagged a gap most frameworks miss: client-supplied AI that arrives in the teleradiologist's worklist with no information on its validation or thresholds.
On LinkedIn, Nina Kottler argued that as AI surfaces more findings, including incidental ones, radiology must help patients understand them, which requires oversight and ongoing governance rather than faster reporting alone. Curt Langlotz shared a Stanford AIMI post on practices building AI in-house, where Woojin Kim supplied the necessary counterweight: the practice still has to show the benefit.
Source articles are indexed in PubMed with verified DOIs. Manuscript-stage work is excluded. The full weekly report is produced in French.
Get the weekly digest
One email a week on what matters in radiology and imaging AI. Researched by a radiologist and imaging informatics professional. Free, no vendor sponsorship.