Radiology AI Snapshot 2025

Highlights at a glance

• MRI and CT dominate research volume, but CT and mixed-modal systems account for a larger share on the FDA side.
• Classification & Segmentation lead academic tasks, while the FDA shows strong presence in Detection and Segmentation.
• Most work remains in-silico—the maturity funnel narrows dramatically by clinical pilots and FDA clearance.
• FDA AI series are accelerating year-over-year, with especially sharp gains from 2023→2025.

Modalities: how research attention is translating

Radiology AI research remains overwhelmingly concentrated in MRI and CT, which together make up more than half of the 6,470 papers in this snapshot. Across the 1,572 FDA-cleared devices we tracked, the mix tilts toward multimodal systems and CT, signaling that vendors prioritize deployable workflows where hardware access and safety cases are already well established. Even within the same modality family, the shape of translation differs: for instance, Mixed Modality is third in publications yet falls behind in regulatory traction, underscoring how high research enthusiasm does not automatically convert into viable devices.

Mixed-modal systems and ultrasound show the fastest proportional gains among cleared devices. Vendors lean on multi-input designs to reduce false positives and to contextualize imaging outputs with EMR or prior studies. That mix-and-match approach complements the steady flow of CT clearances, which remains the workhorse modality because it already fits emergency and inpatient pathways where time-to-diagnosis is critical.

The dual composition chart above highlights how much headroom remains for translation. MRI alone accounts for 1,928 papers, yet fewer than 402 devicesin the same modality have cleared the FDA. Meanwhile, CT maintains the most balanced relationship between academic production and clinical deployment, with the highest absolute number of clearances despite ranking second in publication volume.

Task focus: segmentation vs deployment-ready detection

Academic labs are still captivated by segmentation and classification challenges: the leading three tasks— Classification, Segmentation, and Reconstruction—collectively represent the majority of study endpoints. FDA clearances paint a different picture: Detection, Segmentation, Image Synthesis lead the cleared portfolio, providing quicker clinical wins while bypassing the harder validation burden associated with pixel-perfect segmentation in heterogeneous populations.

The data also show growing enthusiasm for workflow orchestration and reporting assistance tools. These systems bundle AI outputs into existing RIS/PACS, helping radiologists action insights without leaving their reporting environment.

Translation Gap

The largest absolute gaps by modality are MRI (+1,705) and CT (+1,202), which reflects their heavy research investment relative to cleared devices. By task, Classification (+2,602) and Segmentation (+1,250) show substantial research surplus, whereas Triage is an exception with more FDA products than papers (−93), driven by time-critical use-cases (ICH, PE, pneumothorax). The remaining top gaps—Classification (+2,602), Segmentation (+1,250), Reconstruction (+488)—underline how evidence generation slows down as soon as tasks require richer longitudinal follow-up, while modality disparities (MRI (+1,705), CT (+1,202), Mixed Modality (+806)) point to infrastructure and reimbursement hurdles rather than technical readiness alone.

Beyond the headline numbers, ultrasound and PET exhibit smaller but meaningful deltas, hinting at strong vendor momentum in areas where safety cases are easier to articulate (for example, obstetrics and oncology staging). The only negative gap among top tasks is Triage, emphasizing how regulatory approvals chase clinical urgency.

Closing the modality gap will require more prospective evidence and workflow validation in MRI-heavy subspecialties such as neuro and MSK. Sponsors also cite reimbursement alignment and data harmonization as blockers. Targeted public-private registries that pool annotated MRI data would go a long way toward shrinking those deficits.

On the task side, classification still lacks the regulatory clarity that detection enjoys. Many classification systems behave like clinical decision support, blurring lines with practice guidelines. Expect professional societies to issue more guardrails here over the next 12 months as hospitals push for automation that integrates structured reporting.

Maturity of Evidence

The funnel underscores that most work is still in silico (5,424), with comparatively few clinical pilots (400) and very limited FDA-cleared outputs (14 within this study set). Reviews and retrospective studies dominate; prospective and RCTs are rare.

Encouragingly, the proportion of prospective pilots has ticked upward compared with last year, suggesting funders and translational teams are testing real-world deployment earlier in the research lifecycle. Still, the overall funnel conversion remains below 1%, underscoring the need for shared validation infrastructure, post-market surveillance, and multi-site evidence networks.

Each stage of the funnel highlights a different operational hurdle. Concept and prototype work soaks up resources but rarely ships because deployment partners are not lined up early enough. Clinical pilots are the tightest bottleneck: they demand coordination with hospital IT, radiology leadership, and risk management teams, all of which slows the cadence of experimentation. Finally, post-market monitoring remains an untapped differentiator—vendors that invest there can credibly claim safety and effectiveness advantages when marketing to health systems.

Within academia, segmentation of neuro and thoracic imaging stays on top because of rich open datasets and benchmark leaderboards. FDA approvals, in contrast, highlight triage and computer-aided detection products optimized for emergency settings. Vendors have learned that shortening door-to-needle time or flagging worklists delivers the most immediate operational ROI, explaining why triage is the only task with a negative gap (more clearances than papers).

The FDA task mix shows a maturing vendor ecosystem: products balance acute care workflows (stroke, pulmonary embolism) with longitudinal monitoring (oncology follow-up, chronic lung disease). The presence of workflow-support and quality assurance tools among the top five also indicates regulators are comfortable with AI that guides radiologists rather than replaces them.

FDA Momentum

FDA AI/ML device series have grown from 49 in 2018 to 494 in 2025. Annual counts continue to trend upward even when the pace varies year to year, reflecting a broader acceptance of AI tooling in clinical workflows and better productization of detection/segmentation tasks.

Approvals peaked in 2025 with 494 clearances. The cadence has held steady through 2025, and pipeline interviews suggest 2026 will feature more combination devices that pair imaging AI with decision-support dashboards and structured reporting.

Takeaways

Research energy is concentrated in MRI and CT with classification/segmentation tasks. On the translation side, FDA activity emphasizes detection/segmentation, triage for time-sensitive conditions, and multi-modal products. The evidence funnel shows substantial room—and need—for more prospective trials and post-market learning to bridge the gap from promising algorithms to reliable patient impact.

In short, the opportunity space for 2026 and beyond lies in cross-modality orchestration, better evidence generation, and thoughtful integration into radiologist workflows. Teams that invest in measurement, quality, and human-centered design will be best positioned to turn this research momentum into durable, regulated products.

Dig deeper into the underlying records via the research paper library and the FDA clearance tracker, where you can filter modalities, tasks, and source documentation.

Stay ahead of each refresh by subscribing to the Rad AI Slice newsletter.