30-Second Takeaway
- Partially autonomous AI in mammography/DBT can substantially cut reading volume but raises recall rates.
- Commercial DBT AI performs well overall yet underperforms for in-situ disease, calcifications, and dense breasts.
- Head–neck–aortic vessel wall MRI refines supracardiac plaque assessment and reduces ESUS classifications versus CTA.
- Integrating image-based AI with clinical risk factors improves longer-horizon lung cancer risk prediction.
- Unified MRI denoising and CT radiomics can support faster MRI and more consistent nodule assessment but require careful deployment.
Week ending March 21, 2026
AI-enabled imaging and quantitative CT are reshaping cancer screening and thoracic neurovascular practice
Partially autonomous AI workflow boosts mammography/DBT efficiency but increases recall
In this paired noninferiority trial, 31,301 women underwent standard double reading versus a partially autonomous AI-supported screening workflow. The AI strategy auto-cleared exams labeled low risk and double read only higher-risk studies with AI support, reducing radiologist workload by 63.6%. Cancer detection rose 15.2%, from 6.3 to 7.3 per 1,000 screens, but recall increased by 14.8% and failed noninferiority. Workload reductions were similar for digital mammography and tomosynthesis, with detection and recall increases mainly in digital mammography.
DBT AI shows high overall accuracy but notable weaknesses in dense breasts and in-situ disease
This retrospective study evaluated a commercial AI model on 167,860 DBT exams, including 1,368 screen-detected cancers. Overall performance was strong, with AUROC 0.91 and sensitivity 0.73, and was robust across demographic subgroups. Performance declined for in-situ cancers, calcification-dominant lesions, and dense breasts, with AUROCs 0.80–0.88 and lower sensitivities. The model performed best for masses and architectural distortions, with AUROCs around 0.90–0.93 and sensitivities above 0.80.
Head–neck–aortic vessel wall MRI improves plaque detection and clarifies ESUS etiology
A noncontrast neural-network–accelerated head–neck–aortic vessel wall MRI protocol was compared with supra-aortic CTA in 108 stroke or TIA patients. The approximately 15-minute vwMRI detected plaques in 81.5% versus 69.4% with CTA and showed 91% accuracy for calcification versus CTA. vwMRI identified intraplaque hemorrhage in 27.8% of participants and showed similar ulceration detection to CTA. Among 38 CTA-defined ESUS cases, vwMRI reclassified 16, reducing ESUS from 35.2% to 20.4%.
Sybil-Epi improves long-term lung cancer risk prediction beyond LDCT alone
This cohort included 22,469 ever-smokers with 52,482 LDCT series from four screening programs and median seven-year follow-up. The imaging-only Sybil model had AUC 0.93 at one year, declining to 0.79 at six years in independent cohorts. Performance at six years was suboptimal when nodules were absent or small, with AUCs near 0.61–0.64. The new Sybil-Epi model added clinical and epidemiologic variables, improving six-year AUC to 0.83 overall and 0.76 without nodules.
References
Numbered in order of appearance. Click any reference to view details.
Additional Reads
Optional additional studies from this edition.