RSNA 2025: What Radiology AI Got Right This Year

December 8, 2025

Image from RSNA Press Resources

RSNA 2025 delivered fewer buzzwords and more real-world traction for radiology AI. This year’s meeting focused less on speculative prototypes and more on tools that are entering daily use, validated by outcome data, and increasingly embedded in clinical infrastructure. From fully AI-native imaging platforms to validated risk models and FDA-cleared assistive tools, the theme was unmistakable: AI is growing into radiology’s operational layer, not just a diagnostic sidekick.

Below are the core developments from RSNA 2025, distilled for professionals who care about what’s actually being adopted, validated, and regulated.

AI Platforms Are Becoming the Workflow

Major imaging vendors are now building hardware and software ecosystems where AI is deeply integrated at every step of the imaging chain:

GE HealthCare’s Signa One introduced AI that assists in patient positioning via a live camera and adapts scan protocols in real time to breathing motion.
Philips launched Verida, an AI-powered spectral CT system, alongside cardiac MRI automation tools that deliver 14 standardized heart views in under 30 seconds.

Philips Verida CT
Siemens Healthineers unveiled a fully AI-powered imaging chain for interventional systems, embedding AI from acquisition to navigation. It uses real-time anatomical modeling and image fusion to support procedure planning and execution.

What used to be narrow AI tasks like image denoising or segmentation is now baked into the entire imaging process—from acquisition planning to reporting.

FDA-Cleared AI Is Getting Sharper and More Targeted

RSNA also served as the launchpad for several newly FDA-cleared tools with clear clinical applications:

Pristina Recon DL from GE is a deep learning engine for 3D breast tomosynthesis that received FDA premarket approval. In reader preference studies, radiologists favored the AI-reconstructed images in 99% of cases over conventional DBT, citing improved clarity and reduced artifacts without increasing radiation dose.
Syngo.CT Coronary Cockpit from Siemens Healthineers automates segmentation and color coding of coronary plaque on CTA, helping cardiologists visualize disease burden and plan interventions pre-procedure.

Syngo.CT Coronary Cockpit

These tools reflect a maturing ecosystem—one where AI not only works but also passes regulatory scrutiny and delivers measurable value to readers.

Explainability, LLMs, and Responsible AI Took Center Stage

A key theme this year was transparency and responsibility in AI:

Researchers introduced frameworks for clinician-aligned explainability, where AI outputs are structured like clinical reasoning steps rather than simple heatmaps.
Comparisons between general-purpose LLMs and domain-specific models showed better performance for radiology-trained systems when summarizing reports or interpreting modality-specific language.
The FDA emphasized responsible update management via Predetermined Change Control Plans (PCCPs), noting over 350 PCCP submissions to date and 110 AI devices with FDA-approved update plans.

Explainability and governance are no longer academic concerns—they’re part of deployment strategies.

AI Biomarkers Are Becoming Clinically Meaningful

Two standout studies at RSNA 2025 showed how AI-derived image data is crossing into risk stratification and predictive medicine:

Adrenal Volume Index (AVI): Johns Hopkins researchers used AI to quantify adrenal size on chest CTs as a marker of chronic stress. Elevated AVI was linked to higher risk of heart failure and 10-year mortality, offering the first validated imaging biomarker of physiological stress burden.
Clairity Breast: In a 245,000-mammogram study, Clairity’s image-only AI risk model outperformed breast density in predicting five-year cancer risk. High-risk women identified by the AI had a 4.5× higher incidence of cancer, enabling more targeted screening than traditional methods.

These studies demonstrate how radiology AI is evolving from detection tools into personalized risk analytics backed by outcomes data.

The Takeaway for 2025

RSNA 2025 confirmed that radiology AI is no longer about isolated image analysis—it’s becoming foundational infrastructure. Key shifts include:

Imaging platforms built to be AI-native, not just AI-compatible
FDA-cleared tools targeting specific clinical problems with reader-preferred performance
Explainable models that speak clinicians’ language
AI-derived biomarkers with predictive power linked to real outcomes
Regulatory frameworks to support safe AI evolution

This year’s meeting marked a turn toward trust, integration, and practical utility—signaling that radiology AI is entering its next chapter.

If you missed our deeper market scan earlier this year, the Rad AI Snapshot 2025 breaks down the strategic moves, funding trends, and product shifts shaping this evolution.