Prostate MRI AI tested in 22,445 exams

Good morning, there. SCARLET 1 evaluated 22,445 prostate MRI exams across 14 countries.

I see this as a useful stress test for prostate MRI AI beyond single site validation. The site level spread matters because deployment safety will depend on local image quality, protocols, and reader workflow.

How much site level variation would you accept before deploying prostate MRI AI?

Here's what you need to know about Radiology AI last week:

• Prostate MRI AI tested across 22 countries

• Adult CXR AI adapts to neonatal tubes

• 🩻 Mammography AI narrows uncertain reads

• Lung nodule AI boosts clinic throughput

• Plus: 2 newly released datasets, 6 FDA approved devices & 4 new papers.

🧲 Prostate MRI AI tested across 22 countries

RadAISlice: This study stands out for global prostate MRI validation, not just internal performance.

The details:

• 22,445 MRI exams from 21,252 patients informed training and testing

• External testing covered 1,974 patients across 20 cities and 12 countries

• AUCs were 0.86, 0.89, and 0.84 across diagnostic and screening cohorts

• Site variation remained visible, with 28% of sites below AUC 0.80

Key takeaway: For prostate MRI, this moves the conversation from model accuracy toward where AI can safely support readers across sites.

👶 Adult CXR AI adapts to neonatal tubes

RadAISlice: This feels like a practical route for pediatric AI where labeled data remain scarce.

The details:

• 30 NICUs in 20 countries each contributed 50 neonatal chest radiographs

• Continual learning cut tube tip error to 0.67 cm across 29 of 30 sites

• Adult only model error was 4.30 cm, hospital fine tuning was 1.29 cm

• Deep intubation detection improved at 27 of 30 sites

Key takeaway: For pediatric imaging, privacy preserving adaptation may be more practical than waiting for large pooled datasets.

🩻 Mammography AI narrows uncertain reads

RadAISlice: The useful angle here is not standalone AI, but how readers act when uncertainty is explicit.

The details:

• 130,838 mammograms came from 35,709 patients with follow up outcomes

• Unaided AUC was 83.04%, rising to 88.40% in the union scenario

• Sensitivity rose from 54.8% to 83.9% with AI supported uncertainty handling

• Specificity rose from 83.6% to 91.1% in the same scenario

Key takeaway: Mammography AI may add most value when workflow design targets uncertain cases rather than replacing reader judgment.

🫁 Lung nodule AI boosts clinic throughput

RadAISlice: This post implementation study gives a rare look at workflow impact beyond AUC.

The details:

• Three US lung nodule clinics compared matched 12.5 month periods

• New patients rose from 138 to 232 after AI implementation

• Monthly clinic volume increased from 11.0 to 18.6 patients

• Median time to diagnosis stayed 58 versus 57 days, but IQR narrowed

Key takeaway: For nodule programs, AI may expand specialist review capacity, though outcome gains still need stronger prospective evidence.

Synthetic 2D Breast PACT (2026-10-14)

Modality: PACT | Focus: Breast | Task: Image reconstruction; lesion detection

• Size: 3 synthetic datasets. Each has 10,904 2D breast scans; no real patients.

• Annotations: Ground-truth initial pressure, SOS, density, and attenuation maps. Lesion presence, matched lesion-absent controls, and 4 lesion locations.

• Institutions: University of Illinois Urbana-Champaign; Washington University in St. Louis; et al.

• Availability:

  Public. Data repository link not specified. Code: GitHub.

• Highlight: Clinically relevant synthetic breast PACT benchmark with controlled lesions and increasing acoustic modeling complexity.

FDG PET/CT Lung Cancer (2026)

Modality: PET/CT | Focus: Lung, cancer | Task: Classification, explainable AI/localization

• Size: 3,061 2D thoracic slices from 94 patients.

• Annotations: Image-level labels for normal, NSCLC, and SCLC. Cancer subtype labels are pathology-confirmed. No manual segmentations.

• Institutions: Shiraz University of Medical Sciences, Kowsar Hospital

• Availability:

  Public: Zenodo

• Highlight: Public FDG PET/CT lung dataset with pathology-confirmed SCLC/NSCLC labels and patient-level IDs to avoid data leakage.

🏛️ FDA Clearances

• K262015 - Elucid FFR IQ estimates FFR CT from coronary CTA with reported sensitivity 92.2% and specificity 93.1%.

• K262257 - Brainlab spine software adds AI vertebra labeling and planning support across CT, MR, and X ray guided workflows.

• K261225 - CAS One IR supports CT guided interventions with AI segmentation for abdominal organs and lesions.

• K263048 - Siemens Syngo Carbon Space VA4 provides multimodality image display, measurement, and postprocessing tools.

• K261612 - Airs Medical SwiftSight BodyComposition analyzes MRI scans for fat, muscle, and tissue distribution.

• K261042 - GE Lung Guidance uses AI to assist clinicians in acquiring lung ultrasound images.

📄 Fresh Papers

• doi:10.1016/j.acra.2026.08.038 - A prospective pilot found integrated breast US AI shortened exams and raised sensitivity, while sequential AI raised specificity.

• doi:10.1016/j.acra.2026.09.054 - A prospective validation study linked cardiac MRI ECV plus machine learning to 1 year MACE prediction after STEMI.

• doi:10.1007/s00330-026-12970-1 - Prospective PCCT data showed low dose scans preserved deep learning ILD quantification with CTDIvol 0.77 mGy.

• doi:10.1148/rycan.250617 - A five center NSCLC study used CT, radiomics, clinical data, and pathology text to improve recurrence risk stratification.

• Browse 290 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• UC Irvine used NLP on clinical notes and found over three times more lung screening eligible patients than EHR fields.

• Neiman HPI argues imaging demand and radiologist workloads continue to rise despite claims that AI has eased shortages.

• A buyer guide urges radiology teams to scrutinize vendor metrics, validation context, workflow fit, and operating thresholds.

• Radiologists and LLMs showed poor agreement when asked to distinguish real from synthetic chest radiographs.

• Manual and AI segmentations yielded biased radiomics features, underscoring the need to report segmentation methods clearly.

• Read 15 fresh radiology AI news articles for last week.

That's it for today!

👋 Quick favor: drag this into your Primary tab so you don’t miss next week. Or just hit Reply with one thought.

See you next week.

P.S. We keep building free tools to accelerate your radiology work. What's the most time-consuming pain point in your day that we should help speed up? Reply and share your take so we keep building around you.