RSNA 2026 — Tools¶

The RSNA 2026 AI track is dominated by medical-imaging foundation models — self-supervised models pretrained on millions of radiology images (or radiology + radiology-report pairs, or radiology + pathology jointly) that can be fine-tuned or prompted for downstream tasks: detection, segmentation, report generation, triage, risk-stratification, opportunistic screening. This directory captures one page per model / method that surfaces at RSNA 2026 with new checkpoints, clinical-deployment data, or named-trial readouts.

Status: Template stub — meeting is ~7 months out. The per-tool format below is the proposal; tool entries get populated as the program goes live (late summer) and during / after the meeting. Pathology FMs already covered at AACR 2026 (CHIEF, Prov-GigaPath, UNI/UNI2) are cross-linked, not duplicated — see "Cross-vault links" below.

Per-tool entry template¶

Each tool gets its own page (tools/<model-slug>.md) following this structure:

# <ModelName>

**Family:** radiology-fm / pathology-fm / multimodal-fm / segmentation-fm / report-generation / vendor-product
**Modality:** CT / MR / X-ray / mammography / US / multimodal — be specific (e.g., "CT chest only", "whole-slide H&E + CT")
**Released:** YYYY (venue, e.g., "2025 (NEJM AI)", "2024 (Nature)")
**License:** OSI license or proprietary; note non-commercial-academic gating
**Code / checkpoint:** [repo link] + checkpoint availability (open, gated, none public)
**Vendor / sponsor:** academic group, company, or consortium
**Also surfaced in:** AACR 2026 / NeurIPS 2026 / ISMB 2026 / ASCO 2026 — short list of cross-vault hits

> One-paragraph plain-language summary: what the model does, what data it was pretrained on, what tasks it's evaluated for, what's novel about it. ~120 words.

## Architecture & training

2–4 paragraphs:
- Backbone (ViT-L, Swin, hybrid CNN-Transformer, etc.) and parameter count
- Pretraining objective (DINO, MAE, CLIP-style image-text contrastive, MoCo, etc.)
- Pretraining corpus size and composition (#images, #institutions, modalities, anatomies)
- Slide-level / volume-level / report-level aggregation if applicable
- Released checkpoint variants and what differs

## Use at RSNA 2026

<!-- For populated tools: pulled from program / Daily Bulletin / Radiology journal. -->

**Sessions / talks at RSNA 2026 (n=X):**
- Date — session — talk title — speaker — link

**Named clinical trials presenting data at RSNA 2026 (n=X):**
- Trial ID / acronym — sponsor — endpoint — readout (PFS / sensitivity / AUROC / etc.)

**AI Showcase exhibitors using this model (n=X):**
- Company — booth — product

## Cross-vault links

- AACR 2026: [link to the AACR dossier if one exists] — short note on what the AACR side covers
- ASCO 2026: [link if an imaging-driven trial readout appears at ASCO]
- JPM 2026: [link if the sponsor company presented at JPM]
- NeurIPS 2026: [link if a methods paper landed at NeurIPS]
- ISMB / RECOMB 2026: [link if a benchmark paper landed there]

## What's missing / where evidence is weak

Bullet list — honest about gaps: no head-to-head benchmark, license unclear, single-institution evaluation, no prospective deployment data, etc.

## Takeaway

1–2 sentences: what RSNA 2026 actually told us about this model beyond the paper.

Tool index — candidate roster (pre-meeting)¶

Candidate models / methods to watch, grouped by family. Populated entries link out; unpopulated ones are tracked here for the program-notes triage pass. The roster is conservative — only models with a credible RSNA 2026 footprint (named talk, AI Showcase product, companion-publication submission) get individual pages. Otherwise they live in a roll-up paragraph in ../program-notes.md.

Pathology FMs (cross-linked to AACR 2026)¶

These are already covered in the AACR vault — the RSNA page is an extension addendum, not a duplicate.

Model	AACR dossier	RSNA-side notes
CHIEF	aacr-2026/topics/bioinfo-tools/tools/chief.md	Watch for radiology-pathology bridges / multimodal extensions at RSNA
Prov-GigaPath	aacr-2026/topics/bioinfo-tools/tools/prov-gigapath.md	Microsoft + Providence — sponsor may demo at AI Showcase
UNI / UNI2	aacr-2026/topics/bioinfo-tools/tools/uni.md	Mahmood Lab — same lab has CHIEF; watch for joint H&E + radiology presentation

Radiology-native FMs¶

Model	Modality	Sponsor	Status at RSNA 2026	Page
RadFM	multimodal (2D + 3D radiology + text)	Shanghai Jiao Tong + others	candidate — refresher-course material	TBD
Med-PaLM-M	multimodal (radiology + report)	Google DeepMind	candidate — informatics track	TBD
Merlin	abdominal CT + EHR	Stanford	candidate — opportunistic-screening sessions	TBD
RadDino	chest X-ray FM (DINOv2-based)	Microsoft Research	candidate — chest AI sessions	TBD
MAIRA-2	chest X-ray report generation	Microsoft Research	candidate — informatics / governance	TBD
CheXagent	chest X-ray (multimodal LLM)	Stanford AIMI	candidate — informatics	TBD
LLaVA-Med / LLaVA-Rad	vision-language radiology	Microsoft + Stanford	candidate — informatics	TBD

Segmentation FMs¶

Model	Modality	Sponsor	Status at RSNA 2026	Page
MedSAM	universal medical segmentation	Wang lab (UToronto)	candidate — segmentation refresher	TBD
SegVol	3D volumetric segmentation	BAAI / Tsinghua	candidate — segmentation refresher	TBD
Universal-MedSeg variants	various	various	placeholder	TBD

Screening / clinical-deployment AI (named-trial entries)¶

These are models with a credible RSNA 2026 trial readout — entry per model, trial cross-linked in the page.

Model / vendor	Modality	Trial / readout	Page
Mammography AI-CAD (Lunit INSIGHT MMG, ScreenPoint Transpara, Hologic Genius AI, etc.)	mammography	AI-STREAM follow-up, BreastScreen Norway, prospective deployment audits	TBD
Mirai / Mirai-MRI	mammography risk + MR triage	MIRAI-MRI (NCT05968157) — supplemental MRI triggered by AI risk	TBD
Lung-nodule AI (various — Aidoc, Riverain ClearRead, RadAI, etc.)	chest CT	NLST-replay readouts, USPSTF-aligned screening cohorts	TBD
PI-RADS AI (prostate MRI)	prostate MRI	PI-RADS Steering Committee AI-development reporting standards (Radiology v1.0)	TBD
Body-composition / opportunistic-screening AI	abdominal CT / whole-body MRI	RSNA 2026 whole-body MRI body-composition study (May 2026 press release)	TBD

Methods / governance (not models)¶

Topic	Why it matters	Page
AI bias auditing / fairness frameworks	governance refresher courses	TBD
FDA-cleared AI device list / regulatory snapshots	reimbursement + governance	TBD
ACR DSI (Data Science Institute) standards	tied to "Radiology Reimagined" IAIP demo	TBD

Why this format¶

One model per page, AACR-style — the AACR topics/bioinfo-tools/tools/ index uses the same shape. Borrowing it lets a reader hop AACR ↔ RSNA in one click without learning a second schema.
Cross-vault links are first-class — pathology FMs that already have an AACR dossier don't get re-typed; they get a row in the table above plus a back-link in any RSNA addendum.
Trials fold into model pages — at RSNA, most trials are evaluating a model. Indexing by model keeps trial coverage attached to its evidence base.
"Status" tells the reader whether they're looking at a paper-only candidate or a model with named-session footprint — same convention as EuroBioC 2025 tools.

Open questions for the pilot¶

Where do vendor products go? — open: an FDA-cleared AI device that's a wrapper around a public FM should probably link to both (the FM gets the model page; the vendor product gets a row in the "AI Showcase exhibitors" section of that page). Need to validate this against the actual exhibitor list when it goes public.
How granular for screening AI? — open: every major mammography vendor has an AI-CAD product, but lumping them into one "mammography AI-CAD" page hides head-to-head differences. Likely split if any one product has a load-bearing RSNA 2026 readout (e.g., a single-product prospective trial); otherwise stay rolled up.
Cross-modality FMs (radiology + pathology) — open: these straddle AACR and RSNA. Default: page lives in whichever vault hosts the named talk; the other vault gets a back-link. If the model debuts at NeurIPS / arxiv with no conference home, the page goes wherever the first deployment cohort lives.