🩻 Mammography AI risk rose years before diagnosis

Good morning, there. Mammography AI risk scores rose from 2.1 to 6.6 before diagnosis across 158,807 screening mammograms.

I see this as a practical shift from static risk labels to longitudinal imaging biomarkers. If validated prospectively, changing AI scores could inform screening intervals, MRI referral, and prevention discussions.

How would you act on a rising AI risk score before cancer is visible?

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

• 🩻 Breast AI risk changed before diagnosis

• AI improved resident performance on ILD CT

• Radiology kept its lead in FDA AI clearances

• MRI model predicted breast axillary status at scale

• Plus: 7 newly released datasets, 6 FDA approved devices & 4 new papers.

🩻 Breast AI risk changed before diagnosis

RadAISlice: This Radiology study suggests mammography AI risk scores may behave like longitudinal imaging biomarkers.

The details:

• 158,807 screening mammograms from 54,014 women were analyzed

• Cancer cases rose from median risk 2.1 to 6.6 by index exam

• Cancer-free controls stayed near 1.8 to 2.2

• Score slope was 1.13 per year in cancer cases versus 0.09 in controls

Key takeaway: This supports dynamic mammography-based risk tracking as a potential tool for screening interval and prevention discussions, not just one-time risk labels.

🫁 AI improved resident performance on ILD CT

RadAISlice: I see this as one of the clearest workflow-relevant chest CT studies of the week.

The details:

• Multicenter study used 1,097 HRCT exams with 108 external test cases

• AI reached 77.8% accuracy on external testing using MDD reference

• Thoracic experts ranged from 61.1% to 81.5% on the same dataset

• Resident accuracy improved by 14.8 percentage points with AI assistance

• Reading time fell by 20.7% with P below 0.001

Key takeaway: This feels practical for thoracic imaging because it supports less experienced readers on a difficult CT task while improving speed and consistency.

📊 Radiology kept its lead in FDA AI clearances

Image from: Radiology Business

RadAISlice: I read this as a market maturity signal rather than a hype headline.

The details:

• 68 new radiology AI algorithms were cleared in Q1 2026

• Radiology now accounts for 1,163 of 1,524 cleared AI tools

• That equals 76.31% of all FDA-cleared clinical AI

• FDA clearance pace rose from 21 per month in 2024 to about 30 in 2026

Key takeaway: This matters because radiologists remain the main clinical end users of regulated AI, shaping procurement, integration, and postdeployment oversight.

🩺 MRI model predicted breast axillary status at scale

RadAISlice: I like this because it addresses a concrete surgical decision pathway from routine breast MRI.

The details:

• Study included 6,271 breast cancer patients

• Model predicted SLN metastasis and SLN burden plus NSLN disease

• Pooled analysis included 4,081 patients for axillary procedure omission

• Performance was robust across tumor stage, receptor status, and menopausal status

Key takeaway: This could matter for preoperative staging workflows if validated prospectively, especially where imaging may help de-escalate axillary surgery.

BSTT_CT (2026)

Modality: CT | Focus: lung, bone and soft tissue tumors | Task: nodule detection, segmentation

• Size: 61 CT scans from 59 patients; 779 metastatic pulmonary nodules

• Annotations: Pixel-level slice masks for nodules, plus nodule center world coordinates and diameter CSV

• Institutions: West China Hospital, Sichuan University; University of Electronic Science and Technology of China

• Availability:

  public (Zenodo)

• Highlight: Rare lung metastasis dataset from BSTTs with pixel-level annotations. All nodules are malignant and often multiple per patient.

LumbarSR (2026-06-26)

Modality: CT | Focus: lumbar spine, vertebrae | Task: super-resolution, bone microarchitecture analysis

• Size: 30 lumbar vertebral specimens; each has 1 Micro-PCCT scan and 8 paired clinical CT scans

• Annotations: Registered paired volumes and bone mask segmentations in NIfTI; original DICOM also provided

• Institutions: Shanghai Sixth People's Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai Jiao Tong University School of Medicine et al.

• Availability:

  public (Zenodo)

• Highlight: Paired and voxel-wise registered clinical CT with 105 μm photon-counting micro-CT reference across 8 clinical scan settings.

PANTHER (2025-04-15)

Modality: MRI | Focus: pancreas | Task: tumor segmentation, cross-domain segmentation

• Size: 211 scans from 211 patients; 127 diagnostic MRI and 84 MRI-Linac MRI

• Annotations: Expert GTV tumor masks; pancreas masks also provided in training sets. Task 1 test labels use 3-reader STAPLE consensus.

• Institutions: Radboud University Medical Center, Odense University Hospital

• Availability:

  public (Zenodo)

• Highlight: First public benchmark for pancreatic tumor segmentation on both diagnostic MRI and MRI-Linac MRI.

CORTEX (2026-06-25)

Modality: CT | Focus: chest, lungs | Task: VQA, report generation

• Size: 76,177 reasoning traces from 3,039 examinations in 1,304 patients

• Annotations: Four-stage structured reasoning traces with clinician-designed rubric validation; includes task, observation, reasoning, and answer stages

• Institutions: MBZUAI, Hasso Plattner Institute, et al.

• Availability:

  public upon acceptance (code/data page)

• Highlight: Adds validated, clinician-inspired step-by-step reasoning to CT-RATE for trustworthy 3D chest CT MLLMs

BenchX (2026-06-23)

Modality: CT | Focus: pancreas, abdomen | Task: tumor detection, localization

• Size: 85,355 abdominal CT scans from 6 cohorts; patients not fully specified

• Annotations: Scan-level pancreatic tumor presence labels; subset with voxel-wise tumor masks; rich metadata on age, sex, race, CT phase, scanner, spacing

• Institutions: Johns Hopkins University, Stanford University, et al.

• Availability:

  public (planned release); link unspecified

• Highlight: Large benchmark for pancreatic tumor AI with subgroup labels for demographic and CT protocol bias analysis across 6 global cohorts.

MEDLAYXPLAIN-122K (2026-06-19)

Modality: Multimodal (MRI, CT, X-ray, pathology, et al.) | Focus: brain, abdomen | Task: lay captioning, expert-lay alignment evaluation

• Size: 122,789 samples; 79,715 train / 18,484 val / 24,590 test; patient count not specified

• Annotations: ROI grounding with bounding boxes; paired expert and lay captions; UMLS-linked concepts and semantic types

• Institutions: Seoul National University, Seoul National University Hospital, et al.

• Availability:

  public (GitHub); dataset built from public source datasets

• Highlight: First large multimodal benchmark for patient-friendly medical image explanations with region grounding and verified expert-lay caption pairs across 8 modalities

NCCT benchmark (2026-06-15)

Modality: CT | Focus: abdomen, pelvis | Task: disease classification, report generation

• Size: 1,254 scans from 1,254 patients; 1,085 internal and 169 external cases

• Annotations: Paired NCCT volumes with triphasic contrast-enhanced radiology reports; 53 pathology labels extracted from reports and partly radiologist-audited

• Institutions: Ain Shams University, The Hong Kong University of Science and Technology, et al.

• Availability:

  public (code/benchmark)

• Highlight: First multi-center benchmark for multi-organ abdominal diagnosis and contrast-style report generation from non-contrast CT alone.

🏛️ FDA Clearances

• K260406 - Brainomix 360 Hyperdensity is FDA cleared to analyze CT images for brain hyperdensities relevant to acute neuroimaging workflow.

• K253192 - DeepXray Spina is FDA cleared to analyze radiographs for low bone mineral density and support opportunistic osteoporosis assessment.

• K260322 - Acorn 3D Software received FDA clearance for automated image processing and 3D model generation to support planning workflows.

• K250839 - A dental CT X-ray system received FDA clearance for 3D imaging of teeth and surrounding structures for diagnosis and planning.

• K252996 - Konica Minolta Universal 1417PI received FDA clearance as a digital flat panel X-ray imaging device for radiographic acquisition.

• K261352 - 2D Hip Planning Software received FDA clearance to process radiographs for hip procedure planning support.

• Explore last week's 7 radiology AI FDA approvals.

📄 Fresh Papers

• doi:10.3174/ajnr.A9494 - A Medicare analysis found NTAP-billed AI use for suspected LVO peaked at 21% and was tied more to facility factors than patient factors.

• doi:10.1148/ryai.260179 - In the LUNA25 challenge, an AI system outperformed 65 radiologists on malignancy risk estimation for 5 to 15 mm screening lung nodules.

• doi:10.1148/rg.250173 - A Radiographics review outlines human oversight models and monitoring signals for deployed radiology AI.

• doi:10.1111/jebm.70141 - A CLAIM audit of 501 imaging AI papers found median compliance of 51.4%, with gaps in robustness, code sharing, and failure reporting.

• Browse 444 new radiology AI studies from last week.

📰 Everything else in Radiology AI last week

• A new breast MRI method generated dynamic images at one frame per second versus up to two minutes with conventional methods.

• Cardiac imaging leaders report AI is already embedded from acquisition to quantification, often without clinicians realizing it.

• A new framework outlines infrastructure, validation, governance, and AI readiness steps for sustainable digital pathology programs.

• Read 6 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.