AI in Radiology and Medical Imaging: An Evidence-Based Balance Sheet for Deep Learning

Medical imaging is the area where artificial intelligence (AI) has found its fastest and most concrete uptake in healthcare. The reason is straightforward: radiology data is digital, standardized (DICOM), and high-volume. Deep learning models are precisely at home on data like this — extracting patterns from millions of labeled pixels. But "excellent pattern recognition" and "clinical benefit" are not the same thing. The aim of this article is to bridge those two with the evidence.

The Language of Numbers: Radiology's Dominance in FDA Clearances

The U.S. Food and Drug Administration (FDA) regularly publishes a list of AI-enabled medical devices. As of the end of December 2025, the FDA had cleared a total of 1,451 AI-enabled medical devices since 1995; of these, 1,104 (76%) are radiology devices. Radiology has led this list for years; about three-quarters of the new clearances granted in 2025 alone were in radiology.

This concentration is not accidental. Imaging is an ideal setting for machine learning, with a measurable input (the image), a defined output (lesion present/absent, volume, class) and a clear reference standard (pathology, follow-up, expert reading). By company, GE HealthCare holds the most radiology AI clearances (around 120 algorithms), followed by Siemens Healthineers and other major manufacturers. Aidoc, prominent in stroke and emergency triage, has more than 30 FDA clearances; Viz.ai, focused on stroke care coordination, has more than 50 cleared algorithms.

Mammography: AI's Strongest Randomized Evidence

Breast cancer screening is the imaging area in which AI has the most robust chain of evidence. Sweden's MASAI trial (Mammography Screening with Artificial Intelligence) is the first large randomized controlled trial in this space. More than 100,000 women aged 40–80 were randomly assigned either to standard double-reading by two radiologists or to an AI-supported reading arm.

The interim safety analysis published in The Lancet Oncology in 2023 showed that AI-supported reading did not reduce the cancer detection rate and reduced radiologist reading workload by 44%. The trial's final results, published in The Lancet, were stronger still: the AI-supported arm achieved 29% higher cancer detection at screening; sensitivity rose from 73.8% to 80.5% while specificity remained similar (about 98.5%). Moreover, this increase was not merely numerical — the AI arm detected 16% fewer invasive cancers and notably fewer large (advanced-stage) cancers, meaning earlier and more meaningful detection.

Evidence is accumulating at the product level as well. South Korea-based Lunit INSIGHT MMG received FDA 510(k) clearance in 2021 and was trained on more than 240,000 mammography cases (including nearly 50,000 cancer cases). A large population-based retrospective study published in European Radiology in 2023 confirmed that, when the cut-off threshold was matched to first-reader sensitivity, the system's accuracy did not differ in a statistically significant way from radiologists. The system has been reported to offer a relative advantage particularly in dense breast tissue.

MASAI trial, Sweden; The Lancet Oncology (2023, interim analysis) and The Lancet (final results). Values are as reported in the source publications.

Stroke and Pulmonary Embolism: Triage Where Time Is Critical

In acute stroke, "time is brain." Rapid recognition of large-vessel occlusion (LVO) on CT angiography directly affects door-to-needle time for patients eligible for thrombectomy. The Viz.ai platform received FDA clearance for LVO detection with reported high performance (approximately 96% sensitivity, 94% specificity). When the system detects a suspected finding, it alerts the stroke team within seconds via mobile notification, aiming to shorten the delay between imaging and treatment. The same platform also carries cleared modules for other time-sensitive conditions, including pulmonary embolism, aortic disease, and intracranial hemorrhage.

Aidoc, in turn, flags emergency findings on CT — intracranial hemorrhage, pulmonary embolism, pneumothorax, cervical spine fracture — using a "worklist prioritization" logic: it pushes scans containing critical findings to the front of the radiologist's queue. The clinical value here is not "making the diagnosis" but rather getting the right patient to the front at the right time — a safety net and triage layer.

Chest X-ray and Global Health: Tuberculosis Screening

AI's most concrete public-health contribution in resource-limited settings is chest X-ray-based tuberculosis (TB) screening. In its 2021 TB screening guidelines, the World Health Organization (WHO) recommended computer-aided detection (CAD) software as an alternative to a human reader in settings with limited radiologist access. In WHO's evidence appraisal, tools such as qXR, Lunit, and CAD4TB reported area under the curve (AUC) values of approximately 0.76–0.83 in screening use.

Qure.ai's qXR solution, trained on more than a million chest radiographs, has been deployed at thousands of sites in over 60 countries and is among the first AI-enabled chest X-ray tools to receive regulatory clearance. Lunit INSIGHT CXR is designed to detect 10 major chest abnormalities, including nodules, pneumothorax, consolidation, and pleural effusion. The value of these systems lies in their ability to scale screening capacity in areas where specialist radiologists are not available, especially when paired with mobile radiography units.

Limits, Risks, and the Right Expectations

The other side of the picture deserves an honest reckoning. First, distribution shift: a model's performance can degrade if the hospital population, scanner brand, or imaging protocol differs from where it was trained. External (independent) validation is therefore indispensable. Second, automation bias: the temptation for a clinician to over-rely on the AI flag and downgrade their own judgment is a real patient-safety risk. Third, the vast majority of FDA-cleared systems are cleared to assist a radiologist (concurrent/second reader, triage) — not to replace one; the evidentiary bar for fully autonomous reading is much higher and rarely cleared.

Moreover, even though the overwhelming majority of FDA clearances are in radiology, reimbursement infrastructure for these systems still lags in most health systems — meaning there is a gap between technological maturity and financial/operational adoption.

Conclusion

AI in radiology is no longer a "promise of the future" but a tool whose benefit has been proven for specific, well-defined tasks: randomized evidence that AI-assisted mammography screening reduces workload while preserving or increasing detection; triage in stroke and pulmonary embolism that buys time; CAD software that scales TB screening in resource-limited settings. The right frame is to position AI not in place of the radiologist but as a tireless "second eye" that directs attention to the most critical finding — while remembering that every algorithm has its own external validation, its own population, and its own limits.