AI Radiology Accuracy: Prospective Studies and Outcomes

Prospective and externally validated evidence on artificial intelligence (AI) in radiology has matured fastest in breast cancer screening and remains thinner elsewhere, a distinction that matters more than any single accuracy headline. The Swedish MASAI randomized controlled trial, the largest of its kind to date, randomized 105,934 women to AI-supported versus standard double reading and reported sensitivity of 80.5% versus 73.8% with a non-inferior interval cancer rate (^[1]) (^[2]). Independent confirmation comes from Germany's 463,094-woman PRAIM implementation study, which found a statistically significantly higher cancer detection rate with AI support (^[3]), and South Korea's AI-STREAM cohort, which raised detection rate by 13.8% without increasing recall (^[4]).

Outside mammography, evidence is more fragmented and sometimes unfavorable to AI. A prospective, multicenter comparison for intracranial hemorrhage detection found AI-assisted radiologists significantly outperformed standalone AI, with the standalone AI's false-positive count roughly 73 times higher (see Diagnostic Accuracy below), while a prostate MRI validation study found an optimized AI threshold could lower the per-patient false-positive rate below the radiologist's own. Workflow effects are the most consistently positive finding in this literature, with reporting-time reductions ranging from 14.6% in one chest CT nodule study (^[5]) to 63.6% in an autonomous mammography workflow (^[6]), though gains concentrate during high-volume periods and can reverse off-hours (^[7]). Direct patient-outcome data remain the thinnest evidence layer, limited mostly to process metrics such as faster stroke reperfusion and shorter pulmonary-embolism-related hospital stays (see Patient Outcomes and Clinical Impact below).

The FDA's AI-Enabled Medical Device List is intended to identify AI-enabled medical devices authorized for marketing in the United States. Market-sizing estimates vary widely by methodology; one originator with a stated top-down and bottom-up approach sizes the global radiology AI market at $0.61 billion in 2024, growing to $2.27 billion by 2030 (^[8]).

This report separates three evidence layers, diagnostic accuracy, workflow efficiency, and patient outcomes, and treats regulatory authorization as a fourth, distinct question that should never be conflated with the other three. A reader who takes only one finding from this analysis should take this: the strength of AI radiology evidence depends entirely on whether a claim was measured prospectively, externally validated, and against a clinical outcome rather than an accuracy proxy, and each of those three questions can produce a very different answer for the same tool.

Artificial intelligence (AI) has moved from a research curiosity into daily radiology practice, but the strength of evidence behind that shift varies enormously by claim. Regulatory authorization, retrospective benchmark accuracy, and clinical outcome all answer different questions, and conflating them is the most common distortion in vendor and press coverage of the field. The FDA's AI-Enabled Medical Device List is intended to identify AI-enabled medical devices authorized for marketing in the United States.

The FDA states that its AI-Enabled Medical Device List will continue to be updated periodically.

This report instead centers on prospective evidence: studies in which patients were enrolled and imaged going forward in time, ideally with independent or randomized comparison groups, rather than AI models scored retrospectively against archived, often curated, image sets. Prospective and externally validated data are scarcer than retrospective benchmark papers, but they are what determines whether a diagnostic accuracy signal survives contact with routine clinical workflow. The sections that follow separate three distinct evidence layers: diagnostic accuracy under prospective or randomized conditions, workflow and radiologist-performance effects, and patient-level outcomes, before turning to the regulatory landscape, a consolidated data section, real-world deployment examples, and implications for organizations evaluating these tools.

This report synthesizes findings from peer-reviewed prospective cohort studies, randomized controlled trials (RCTs), externally validated diagnostic-accuracy studies, systematic reviews, and official regulatory sources published or updated through September 2026. Every quantitative claim is traced to its originating publication, filing, or regulator, not to a secondary summary. Studies were included if they met at least one of the following criteria: (1) prospective enrollment, meaning patients or images were collected going forward under a predefined protocol rather than pulled retrospectively from an archive; (2) external validation, meaning the AI model was tested on data from a site, population, or time period distinct from its training set; or (3) randomization, meaning patients or readers were formally assigned to AI-assisted versus standard-of-care arms.

A broader systematic review of AI generalizability similarly found specificity losses of up to roughly ~24 percentage points when models moved from internal to external validation, with one example AUC falling from 0.95 internally to 0.91 externally (^[9]). External validation is formally defined in this literature as testing on an independent dataset from a different site or population than the one used for development.

Reporting quality is itself an active methodological concern. The STARD 2015 checklist for diagnostic accuracy studies, radiology's long-standing reporting standard, does not address the issues and challenges raised by artificial intelligence (^[10]), which motivated the STARD-AI extension, an AI-specific version of the STARD checklist focused on reporting AI diagnostic test accuracy studies, finalized in Nature Medicine in 2025 (^[11]) with input from more than 240 international stakeholders, and the full checklist is maintained publicly by the EQUATOR Network (^[12]). A parallel guideline, TRIPOD+AI, updated the older TRIPOD standard to harmonize the landscape of prediction model studies regardless of whether regression or machine-learning methods were used (^[13]), warning that prediction-model studies are frequently poorly conducted and incompletely reported, leaving them at high risk of bias. This report favors studies that at least approximate these standards and flags vendor involvement or funding wherever a study discloses it.

Breast Cancer Screening

Mammography carries the largest and most mature body of prospective AI evidence of any imaging modality, anchored by MASAI (Mammography Screening with Artificial Intelligence), a Swedish randomized controlled trial. MASAI randomized 105,934 women 1:1 to AI-supported double reading or standard human double reading (^[1]), enrolling at four Swedish sites between April 2021 and December 2022 and is registered as completed under NCT04838756, sponsored by the academic body Region Skane rather than an AI vendor (^[14]). Its final results, published in the Lancet in January 2026, reported sensitivity of 80.5% for AI-supported reading versus 73.8% for standard double reading, with an identical, non-inferior interval cancer rate: a proportion ratio of 0.88 between arms (^[2]) (^[15]). A press release accompanying the final trial reported the raw interval cancer counts as 82 of 53,043 women in the AI arm versus 93 of 52,872 in the control arm over two years of follow-up, with 81% of screening-detected cancers caught at screening in the AI arm versus 74% in the control arm, and 16% fewer invasive interval cancers in the AI-supported group (^[16]) (^[17]). An earlier MASAI safety interim analysis, published in Lancet Oncology in 2023 with 80,033 women enrolled, had already shown AI-supported reading cut radiologists' screen-reading workload by 44% (^[18]), with recall rates of 2.2% in the AI arm versus 2.0% in the control arm (^[19]). Oncology trade press has described MASAI as the first randomized controlled trial investigating AI in breast cancer screening and the largest to date (^[20]).

Independent confirmation comes from PRAIM, a German nationwide observational implementation study that compared AI-supported double reading against standard double reading under real-world conditions across 12 sites, screening 463,094 women (260,739 with AI support) read by 119 radiologists between July 2021 and February 2023, and found the AI-supported group's cancer detection rate of 6.7 per 1,000 women was statistically significantly higher than the control group's (^[3]) with a non-inferior recall rate. In South Korea, the AI-STREAM prospective multicenter cohort study, registered on ClinicalTrials.gov as NCT05024591 with sponsorship from an academic hospital and a Korean government research institute rather than a private AI vendor (^[21]), enrolled 24,543 women within the national screening program, recording 140 screen-detected cancers, and found AI-assisted single reading raised cancer detection rate by 13.8% without a significant change in recall rate (^[4]). A subsequent secondary analysis of the AI-STREAM cohort reported the underlying computer-aided detection (CAD) tool's standalone performance at 89.9% sensitivity, 94.3% specificity, and an 8.7% positive predictive value (PPV, the share of positive calls that are true positives) for recall (^[22]).

A Spanish prospective, paired non-inferiority trial of 31,301 women, published in Nature Medicine in March 2026, evaluated a more autonomous AI workflow (where AI independently clears a subset of normal-appearing exams) and found a 63.6% reduction in radiologist workload alongside a 15.2% higher cancer detection rate (rising from 6.3 to 7.3 per 1,000), though its recall rate did not meet the pre-specified non-inferiority margin: it was 14.8% higher than standard reading (^[6]). Notably, that trial disclosed no industry funding, though one co-author is an employee of the AI vendor whose product was evaluated (^[23]). Separately, a multireader crossover study of 14 radiologists interpreting 65 mammograms found AI assistance sharply improved inter-rater agreement on breast-density scoring, with kappa values (a statistical measure of agreement beyond chance) rising from 50.01% to 81.38% (^[24]), suggesting a consistency benefit distinct from detection accuracy.

Chest, Thoracic, and Cross-Sectional Imaging

Outside breast screening, prospective evidence is more fragmented but growing. A single-center, open-label RCT in Korea (n=911) tested AI-assisted low-dose chest computed tomography (CT) nodule evaluation and found no statistically significant change in interpretation time (187 versus 172 seconds; p=.23) but a significantly higher detection rate of Lung-RADS-positive (clinically actionable) nodules: 16.9% with AI assistance versus 10.3% without (^[25]) (^[26]), with overall nodule detection roughly 1.6 times higher (52.9% versus 32.6%) and no lung cancers diagnosed in either arm over roughly 215 days of follow-up (^[27]). A separate prospective AJR study of an AI triage system for incidental pulmonary embolism (IPE) found radiologists reading without AI assistance had significantly lower sensitivity than radiologists reading with AI, 80.0% versus 96.2%, without a significant difference in specificity, which was 99.9% in both arms (^[28]) (^[29]); report turnaround time for positive cases was not significantly different between arms (^[30]).

A prospective, multicenter, real-world deployment of an AI chest-radiograph reporting model (Janus-Pro-CXR, fine-tuned from a large language model architecture) enrolled 296 patients across three Chinese hospitals and found AI-assisted junior-radiologist reports scored significantly higher on quality (mean 4.36 versus 4.12) while cutting interpretation time by 18.3% relative to standard practice (^[31]) (^[32]) (^[33]); its underlying retrospective validation reported AUC values above 0.8 for six key chest findings, including 0.931 for pleural effusion detection (^[34]).

Tuberculosis (TB) screening on chest radiographs is a further prospective use case, largely in global health settings. A five-country prospective cohort study (n=1,392) found CAD4TB version 7 achieved higher specificity (70.3%) than a molecular test (Xpert HR, at 65.1%) when both were fixed at 90% sensitivity (^[35]). A separate prospective diagnostic accuracy study in Oromia, Ethiopia (n=478) found CAD4TB achieved 0.77 sensitivity and 0.93 specificity against sputum testing as the reference standard (^[36]). A large population-scale external validation of a different TB model (AIRIS-TB) on over one million chest X-rays reported an AUC of 98.51% and a lower false-negative rate than radiologists (1.57% versus 1.85%) (^[37]). These results should be interpreted separately: sensitivity is threshold-specific, whereas AUC summarizes discrimination across thresholds, so they are not AUC-equivalent measures or a direct numerical comparison.

Neurological and Other Modalities

Outside chest and breast imaging, prospective evidence is thinner and more mixed. A prospective, multicenter diagnostic accuracy study across 67 medical organizations in Moscow (April 2022 to December 2024), analyzing 3,409 brain CT studies including 1,101 intracranial hemorrhage (ICH) cases, directly compared three standalone commercial AI services against radiologists who had access to those same AI results as an auxiliary tool, and found the AI-assisted radiologists significantly outperformed standalone AI on every core metric: sensitivity of 98.91% versus 95.91%, specificity of 99.83% versus 87.35%, and overall accuracy of 99.53% versus 90.11% (^[38]). The gap was driven almost entirely by false positives: standalone AI flagged 293 false-positive cases against 4 for the AI-assisted radiologists, a roughly 73-fold difference, and diagnostic errors were largely complementary, with all 12 cases missed by radiologists caught by AI and all 45 cases missed by AI caught by radiologists, consistent with a "second reader" implementation model rather than autonomous standalone deployment. The same study's authors note that in the wider ICH literature, positive predictive value fluctuates from 56.7% under multicenter validation conditions to 98.4% in single-center studies, underlining how site-dependent AI performance figures can be.

In prostate imaging, a prospective validation study of an AI tool for detecting clinically significant prostate cancer (csPCa) on biparametric magnetic resonance imaging (MRI) found the software, at an optimized detection threshold, achieved higher specificity than the radiologist (0.68 versus 0.57) with a lower average false-positive rate per patient (0.33 versus 0.41) (^[39]). The study's authors caution that the tool's originally deployed threshold had deliberately prioritized sensitivity and accepted a higher false-positive rate to minimize the risk of missed cancer, meaning the more favorable false-positive figures reflect a retrospectively optimized threshold that itself awaits prospective confirmation.

Workflow effects are among the most consistently positive findings in this literature, though their magnitude is highly context-dependent. A prospective observational study of 11 radiologists reading 18,680 chest radiographs found total reading times were significantly shorter with AI assistance than without, 13.3 seconds versus 14.8 seconds (^[40]), but that time saving vanished when the AI itself flagged an abnormality requiring closer radiologist review (18.6 versus 18.4 seconds). A retrospective real-world study of 19,433 patients (39,323 chest CT exams) at a Dutch academic center found commercial pulmonary-nodule AI reduced adjusted median reporting time by 14.6%, from 21.3 to 18.2 minutes (^[5]), with exploratory modeling suggesting the tool's time savings equated to roughly half a full-time-equivalent radiologist at typical institutional volumes (^[41]). That benefit was uneven across subgroups: thoracic-subspecialty radiologists saw the largest gains, while emergency department exams actually took 7.1% longer under the AI-assisted workflow (^[42]).

Turnaround-time effects follow a similar workload-dependent pattern. An FDA and University of Chicago collaborative study of 11,252 CT pulmonary angiography (CTPA) scans found mean turnaround time for PE-positive examinations during work hours was 68.9 minutes before AI triage and 46.7 minutes after AI triage; the observed work-hour saving was significant, whereas the off-hour saving was not (^[7]). A separate study of 2,501 CTPA exams found an AI worklist-reprioritization tool significantly shortened mean turnaround time (47.6 versus 59.9 minutes) and wait time (21.4 versus 33.4 minutes) for pulmonary-embolism-positive cases without a significant difference in mean read time; the tool reprioritized 12.7% of post-AI examinations (^[43]).

Adoption survey data show usage climbing steadily but from a modest base and with real skepticism among non-adopters. The American College of Radiology's inaugural 2020 survey of 1,427 radiologists found roughly 30% used AI in clinical practice, with 80% of non-adopters citing that they "see no benefit" as their reason for not adopting (^[44]) (^[45]). By early 2024, a European Society of Radiology and EuSoMII survey of 572 members found adoption had nearly doubled to 47.9%, with a further 25.3% of non-users planning to adopt, and computed tomography (38.8% of mentions) and radiography (24.0%) as the most commonly mentioned modalities in that survey (^[46]). Even so, industry surveys on AI adoption and return on investment have found that adopting institutions themselves are often less confident about the technology's uptake than the vendors promoting it, citing concerns about trust, transparency, and cost, a reminder that measured efficiency gains and organizational confidence do not always move together.

Diagnostic accuracy and workflow speed are necessary but not sufficient evidence of patient benefit; relatively few studies follow patients far enough to measure outcomes rather than proxies. MASAI remains the strongest available outcome signal in breast screening because interval cancer rate, the rate at which cancers surface clinically between scheduled screenings, functions as a safety outcome tied to missed or delayed diagnosis. Its non-inferior interval cancer rate combined with a higher screening-detected share and fewer aggressive-subtype interval cancers in the AI arm (^[47]) is consistent with, though not proof of, an eventual stage-shift and mortality benefit; neither MASAI nor PRAIM has yet reported mortality data, since breast-cancer mortality differences typically require a decade or more of follow-up to detect.

Stroke and pulmonary embolism triage provide a different outcome window: minutes matter directly for tissue survival. A single-center before-and-after study of an AI large-vessel-occlusion (LVO) triage tool found it significantly reduced door-to-endovascular-thrombectomy (EVT) time by 30.2 minutes (^[48]), and functional outcome (measured by the modified Rankin Scale, or mRS, where 0 to 1 indicates minimal or no disability) improved numerically from 26% to 40% good outcomes, though this difference did not reach statistical significance in the study's sample size. A second, separate before-and-after study of a different AI stroke workflow tool found door-in-to-puncture time fell from 206.6 to 119.9 minutes and successful reperfusion (a measure of restored blood flow, graded modified TICI 2B to 3) rose from 84.9% to 94.1%, a statistically significant improvement, while in-hospital mortality showed no significant difference between periods (^[49]). Both are single-center, non-randomized, before-and-after designs, so the observed process-metric gains are more secure than any causal claim about mortality.

A comparable pattern appears for pulmonary embolism (PE). A retrospective observational study of an AI-guided pulmonary embolism response team found hospital length of stay fell from 33.6 hours pre-AI to roughly 13 to 19 hours across two post-implementation years, alongside a sharp rise in catheter-based thrombectomy use (^[50]), but the study's own authors note it did not measure mortality or time-to-diagnosis, limiting how far the length-of-stay finding can be extended into a survival claim. Taken together, the outcome literature currently supports confident claims about process metrics (workload, turnaround time, detection rate, procedural speed) and only cautious, hypothesis-generating claims about hard clinical endpoints such as mortality, which remain the least studied layer of evidence in this field.

Regulatory status is frequently the least understood layer of this evidence chain. The FDA's own framing is explicit: its AI-Enabled Medical Device List exists to identify AI-enabled medical devices that are authorized for marketing in the United States (^[51]), following a focused review of the device's overall safety and effectiveness (^[52]), not a determination of comparative clinical benefit.

Market-sizing estimates for AI in radiology vary widely across research firms and are frequently cited without a stated methodology, a common problem when comparing figures across vendors. One originator that publishes an explicit top-down and bottom-up methodology with primary interviews and data triangulation, MarketsandMarkets, sizes the global radiology AI market at $0.61 billion in base year 2024, projected to reach $2.27 billion by 2030 at a 24.5% compound annual growth rate (^[8]) (^[53]). Readers should treat any radiology AI market-size figure as methodology-dependent rather than a settled fact, given how differently research firms scope "the market."

Table 1 below summarizes selected prospective and randomized AI radiology studies discussed in this report, spanning breast, chest, neurological, and prostate imaging.

This table underscores a consistent pattern: prospective and randomized breast-screening trials show the most mature, convergent evidence (higher sensitivity or detection rate with workload relief), while single-study results in chest, neurological, and prostate imaging show larger, less consistent effect sizes and, in the ICH case, a clear net disadvantage for standalone AI relative to human readers.

Systematic reviews provide the clearest picture of how AI diagnostic accuracy varies once results are pooled across many individual studies rather than read one paper at a time. A large systematic review and meta-analysis of deep-learning diagnostic accuracy in medical imaging, covering multiple modalities, found pooled AUC for breast cancer detection ranged from 0.868 (MRI) to 0.909 (ultrasound), with mammography at 0.873, though it noted no breast-imaging studies in the pooled set used prospectively collected data (^[55]), an important caveat given the strong prospective breast-screening trials summarized above post-date that review. For chest imaging, the same review found pooled accuracy varied enormously by task: pneumothorax detection showed relatively low pooled sensitivity of 0.70 alongside high specificity of 0.94, while tuberculosis detection on the same chest-radiograph modality achieved much higher and more balanced pooled sensitivity of 0.95 and specificity of 0.97, and only two chest studies in the entire pooled set used prospective data. The review's authors caution that, due to high heterogeneity and variance between studies, there is considerable uncertainty around estimates of diagnostic accuracy, a caution this report extends to every single-study figure cited above.

False positives, specifically, receive less consistent attention than sensitivity in the underlying literature. A rapid scoping review of AI diagnostics in radiology practice found that of 25 studies measuring sensitivity, 19 reported improvements with AI, 5 reported no change, and only 1 reported a reduction, while specificity results were far less consistent: of 23 studies measuring specificity, only 13 reported improvements, 7 reported no change, and 3 reported an outright reduction (^[56]). The reviewers explicitly flag the risk of increasing false positives, and the wider impact of AI on workflow efficiency, as an unresolved evidence gap.

Table 2 below consolidates the sensitivity, specificity, and false-positive figures discussed across this report's core studies, to make the accuracy-versus-false-positive trade-off directly comparable across tasks.

The table makes visible what individual abstracts often obscure: AI systems that raise sensitivity frequently do so by accepting more false positives, and the size of that trade-off differs by an order of magnitude between tasks, from a negligible specificity cost in mammography to a 73-fold false-positive increase in one intracranial hemorrhage comparison. No single aggregate "AI radiology accuracy" figure can responsibly summarize this table; the task, population, and threshold setting each materially change the answer.

Several of the prospective studies cited above are themselves real-world deployments rather than controlled experiments, and are worth highlighting as operational examples distinct from their accuracy findings. The Janus-Pro-CXR rollout across three Chinese hospitals is a rare example of a foundation-model-based reporting tool measured prospectively in live clinical use rather than only on a retrospective test set, with real-world deployment involving 296 patients and significantly improved report quality scores relative to standard care (^[31]).

The AI-guided pulmonary embolism response team program is a comparable operational case study: an institution restructured its emergency PE response process around AI-flagged triage across a pre-AI period and two full years of post-implementation data, tracking length of stay and thrombectomy utilization (see Patient Outcomes and Clinical Impact above) as concrete operational outcomes rather than only diagnostic accuracy. Similarly, the two independent stroke-center deployments summarized in the outcomes section, one using the Heuron ELVO tool and one using a Viz-branded LVO workflow tool, each represent a full before-and-after institutional workflow change rather than a laboratory benchmark, illustrating how AI triage tools are increasingly evaluated as part of an entire care pathway (door arrival to reperfusion) rather than as an isolated image-reading step. Each of these deployments is single-center or limited-multicenter and non-randomized, so they should be read as operational proof-of-concept evidence rather than confirmatory trials; the field currently has more of this kind of implementation case study than it has large multicenter randomized outcome trials outside of mammography.

The evidence assembled here points toward several practical implications for radiology departments, health systems, and life-sciences organizations assessing AI diagnostic tools. First, procurement and clinical-governance decisions should separate three distinct questions that vendors often blend into one pitch: what is this tool's FDA marketing-authorization pathway and intended use, has its accuracy been externally validated on data resembling the deploying institution's population, and does it have prospective evidence of workflow or outcome benefit in a comparable setting. As this report has shown, a device can clear the first bar with only retrospective data and still see meaningful accuracy erosion in prospective, external, or off-label use.

Second, the workflow evidence suggests AI's time and turnaround benefits concentrate during high-volume, routine periods and can disappear, or even reverse, during off-hours or in subspecialty-mismatched settings, meaning institutions should pilot tools against their own case mix and staffing patterns rather than relying on published averages from a different practice environment. Third, the intracranial hemorrhage and prostate-MRI false-positive findings argue for AI implementations that keep a human reader as the final arbiter (a "second reader" model) in categories where standalone AI specificity lags meaningfully behind trained radiologists, rather than autonomous triage, until larger prospective trials narrow that gap.

For life-sciences and technology organizations advising on these deployments, understanding this evidence hierarchy, rather than a single reported accuracy number, is a precondition for sound governance recommendations. Looking forward, the reporting-guideline efforts (STARD-AI, TRIPOD+AI) and the growing number of multicenter prospective registries suggest the next two to three years should produce a denser prospective evidence base, particularly outside mammography, where randomized designs remain rare.

What counts as a "prospective" AI radiology study? A prospective study enrolls patients or collects images going forward under a predefined protocol, rather than pulling archived images retrospectively; this report also treats external validation, testing on data from a different site or population than the training set, as a related, complementary form of rigor (see Methodology and Study Inclusion Criteria).

What does "AI radiology clinical trial results" actually show so far? The strongest randomized results come from breast screening, where MASAI found higher sensitivity (80.5% versus 73.8%) with a non-inferior interval cancer rate across 105,934 women (^[2]); outside mammography, most available "trial" evidence is smaller, single-center, or non-randomized.

How is AI diagnostic accuracy typically measured on imaging studies? Studies report sensitivity, specificity, positive predictive value, and AUC against a defined reference standard (such as pathology-confirmed cancer, expert consensus, or a molecular test); as the meta-analysis discussed in Data Analysis and Evidence shows, pooled figures range from 0.868 to 0.909 AUC for breast imaging modalities and from 0.70 to 0.95 pooled sensitivity depending on the chest-imaging task.

Does AI improve radiology workflow efficiency in practice? Frequently yes for reporting and turnaround time, most clearly during high-volume periods, with reductions ranging from about 14.6% to 63.6% in cited studies, but effects can vanish or reverse in off-hours or emergency settings (^[5]).

Is there evidence AI radiology tools improve patient outcomes, not just accuracy? Direct patient-outcome evidence (as opposed to accuracy or workflow proxies) is the thinnest layer of this literature; the clearest signals are process-level, such as faster stroke reperfusion times and shorter pulmonary-embolism-related hospital stays, with mortality and long-term outcome data still largely unavailable, as detailed in Patient Outcomes and Clinical Impact above.

What does the FDA AI-Enabled Medical Device List identify? The FDA's AI-Enabled Medical Device List identifies AI-enabled medical devices authorized for marketing in the United States.

What is the false-positive rate for AI radiology tools? It varies enormously by task and cannot be summarized as a single number; the intracranial hemorrhage and prostate MRI comparisons discussed above (see Table 2) show standalone AI false-positive rates ranging from roughly 73 times the radiologist rate in one brain CT comparison to a lower per-patient rate than the radiologist in one retrospectively optimized prostate MRI threshold.

How widely has machine learning radiology adoption spread among radiologists? Survey-based adoption climbed from roughly 30% of respondents in a 2020 U.S. survey to 47.9% in a 2024 European survey, as detailed in Workflow Efficiency and Radiologist Performance above, though these are different populations and not a single continuous trend line.

Prospective and externally validated evidence on AI in radiology is more mature in breast cancer screening than anywhere else in the specialty, anchored by a large randomized trial (MASAI, 105,934 women) and two large independent cohorts (PRAIM and AI-STREAM) that converge on higher sensitivity or detection rate without a corresponding rise in recall rate. Outside mammography, prospective evidence remains fragmented across chest, neurological, and prostate imaging, with effect sizes and even the direction of the accuracy trade-off varying by task, population, and validation setting. Workflow benefits, particularly reduced reporting and turnaround time, are the most consistently documented gains, but they concentrate during high-volume periods and can reverse in off-hours or subspecialty-mismatched use. Direct patient-outcome evidence, as distinct from accuracy or workflow proxies, remains the thinnest and least mature layer of this literature. The FDA's AI-Enabled Medical Device List identifies AI-enabled medical devices authorized for marketing in the United States. Readers evaluating any single accuracy claim in this field should ask three questions this report has organized around: was the evidence prospective or retrospective, was it externally validated, and does it measure accuracy, workflow, or a hard clinical outcome.