필사 모드: AI in Medical Imaging and Radiology 2026 — Aidoc · Lunit · Annalise.ai · Qure.ai · Viz.ai · Tempus · PathAI · Paige · Mediwhale Deep Dive

Prologue — Why medical imaging AI, why now

Spring 2026, an emergency department in a large US hospital. 3:14am, a man in his sixties is brought in with headache and vomiting. The triage nurse assigns NIHSS 4, the ED physician orders a non-contrast head CT. Three minutes later, before the patient is even off the gantry, Aidoc auto-detects intracranial hemorrhage (ICH) and a red banner appears on the radiology console. The same study is also routed to a Viz.ai server, which evaluates large-vessel occlusion (LVO). If LVO is positive, the neurointerventional team's phones light up at the same moment. By the time the on-call radiology resident opens the case, the neurosurgery attending and the interventional team have already seen the images.

A decade ago this chain took 30 to 60 minutes. Radiology read queue, call, fax, call again. The 2026 standard is "from CT gantry to neurointerventional team page in under five minutes." That five-minute delta moves a single patient's mTICI score, and 90 days later, their mRS.

At the same moment, in a mammography suite at Gangnam Severance in Seoul, Lunit INSIGHT MMG serves as a radiology resident's second eye. In an endoscopy suite at the University of Tokyo Hospital, AI Medical Service's GI Genius highlights polyps in real time. At a tuberculosis screening camp outside Mumbai, Qure.ai's qXR rides along with a portable X-ray unit, village by village.

This post is the spring 2026 map of the medical imaging AI ecosystem behind all of those scenes. Over 950 FDA approvals, companies in the US, Korea, Japan, India, Israel and Australia, across radiology, pathology, ophthalmology, cardiology, dermatology, surgery, genomics and clinical LLMs — in a single read.

Chapter 1 · Why imaging became AI's first clinical beachhead — three forces

It is no accident that medical imaging is where AI first landed in clinical practice. Three pressures lined up at the same time.

**First, the radiologist shortage.** The American College of Radiology's 2024 workforce report notes that a US radiologist reads on average about 14,000 studies a year. If a CT takes eight minutes to read, an eight-hour day caps out around 60 — yet queues commonly carry twice that. Teleradiology firms (vRad, Nighthawk, Radiology Partners) plug the gap with globally distributed reads, but demand has grown faster.

**Second, screening at scale.** Mammography, lung CT, colonoscopy, retinal photography — periodic screening of asymptomatic populations has pushed the volume of images to be read up by an order of magnitude. Korea's national screening recommends biennial mammograms for every woman aged 40 and over. Japanese gastric cancer screening and US LDCT lung cancer screening (annual for ages 50+ with smoking history) generate the same pressure.

**Third, quantifying triage.** Clinical evidence that "minutes to critical finding" drives outcomes (stroke, MI, aortic dissection) has accumulated, and insurers, hospitals and regulators have all started to require shorter turnaround time (TAT). US CMS applied NTAP (New Technology Add-on Payment) to LVO triage AI; the UK NHS funds chest imaging AI from a separate budget.

The market built by these three forces is around USD 3.5 billion in 2026 (multiple vendor estimates) and growing more than 30% a year.

Chapter 2 · FDA-cleared AI/ML devices — the era of 950+

The FDA publishes a list of AI/ML-enabled medical devices. It stood at 521 at the end of 2022, passed roughly 880 in 2024, and by the end of 2025 it has crossed 950. The list grows by three to four new approvals every week on average.

The distribution is sharply skewed. **Radiology is about 76%** — dominant by a wide margin. Cardiology is next at around 10%, with neurology, ophthalmology and pathology each at 2–3%. This is not because medical imaging is reducible to pixel pattern matching; it is because DICOM was standardized early, picture archiving (PACS) was mandated, and labels are relatively well defined. Tagging a chest X-ray with "nodule yes/no" is comparatively tractable.

The dominant clearance route is **510(k)** — show "substantial equivalence" to an existing predicate device. New categories (pulmonary embolism triage when it first appeared) take the **De Novo** path, and follow-on devices then flow through 510(k). **PMA (Premarket Approval)** is reserved for a small set — devices that issue an autonomous diagnosis (IDx-DR and a few others).

In 2024 the **PCCP (Predetermined Change Control Plan)** guidance was finalized. Models can now be retrained on a regular cadence without requiring a new 510(k) every time, provided the scope of change and the validation method are agreed with the FDA in advance. PCCP is the structural turning point of the 2026 medical AI market — a model can finally be a living asset.

Chapter 3 · Aidoc — the de facto triage AI standard

Tel Aviv-based **Aidoc** was founded in 2016. The founders — Elad Walach, Michael Braginsky and Guy Reiner — came out of Israel's Unit 8200 imaging community. The first product, head CT **intracranial hemorrhage triage**, got FDA clearance in 2018.

By 2026 the Aidoc portfolio has about fourteen algorithms.

- **BriefCase ICH** — intracranial hemorrhage (SDH, EDH, IPH, SAH)

- **BriefCase LVO** — large-vessel occlusion

- **BriefCase PE** — pulmonary embolism

- **BriefCase C-Spine** — cervical spine fracture

- **BriefCase Aortic Dissection** — aortic dissection

- **BriefCase Rib Fracture** — rib fracture

- **BriefCase Pneumothorax** — pneumothorax

- **BriefCase Free Gas** — intra-abdominal free gas

- **BriefCase Incidental PE / Incidental ICH** — incidentals

- five more

All algorithms run on the same platform — **aiOS** (the Aidoc operating system). When studies arrive from the hospital PACS (Sectra, Philips IntelliSpace, GE Centricity), aiOS runs every algorithm in parallel and bumps positive patients to the top of the worklist. For the radiologist, this is a single red banner that says "see this patient first."

Aidoc's edge is **workflow integration**. A small Israeli startup got into roughly 1,000 US hospitals not on accuracy alone but on a "turn it on in five minutes without touching PACS, RIS or EMR" integration story. In 2024 the NVIDIA partnership standardized inference on MONAI; in 2025 the SMART on FHIR coupling with Epic deepened.

Pricing runs in the low-to-mid six figures USD per hospital per year, and for NTAP algorithms (LVO, PE) part of the cost is recoverable through Medicare.

Chapter 4 · Lunit — Korea's medical AI leader

Seoul-based **Lunit**, headquartered in Gangnam, was founded in 2013. CEO Brandon Suh (Baek Seung-wook) is a KAIST alumnus, and the company has applied a research-first culture in the DeepMind mold to medical imaging. After a 2022 KOSDAQ listing, the market cap has hovered around KRW 2 trillion.

The flagship products are two.

**Lunit INSIGHT CXR** — chest X-ray. It surfaces ten core findings (nodule, consolidation, pneumothorax, pleural effusion, pneumonia, atelectasis, tuberculosis, mass, cardiomegaly, and more). It received a US FDA 510(k) in 2024 and holds approvals across roughly 40 countries (Korea MFDS, Japan PMDA, EU CE, Australia TGA, Brazil ANVISA). Validation papers appear in *Radiology* and *The Lancet Digital Health*.

**Lunit INSIGHT MMG** — digital mammography (2D). Works across Hologic, GE, Siemens and FujiFilm units. The Swedish MASAI RCT published in 2023 in *European Radiology* reported a roughly 20% higher cancer detection rate in the arm that used INSIGHT MMG. That finding has driven adoption into European screening programs.

Since 2024 the focus has shifted to **INSIGHT DBT** (digital breast tomosynthesis). DBT carries tens of times more data than 2D mammography (50–80 slices), and AI directly attacks the reading-time burden.

Another arm is **Lunit SCOPE** — pathology slide quantification of tumor-infiltrating lymphocytes (TIL). SCOPE drives response prediction in immuno-oncology, and pharma companies such as AstraZeneca and Roche have adopted it for trials. This is an interesting path where imaging crosses from diagnosis into therapy selection.

Chapter 5 · Annalise.ai — Australian chest X-ray precision

Sydney-based **Annalise.ai** started as a clinician consortium. The parent company Harrison.ai was built by the Aengus and Dimitry Tran brothers (radiologist and computer scientist respectively). Their vision was simple — "surface every finding a radiologist would look for, on every chest X-ray, at the same time."

The headline product **Annalise CXR** scores **124 findings** simultaneously. Beyond simple nodules and pneumothorax, it covers vertebral compression fractures, diaphragmatic position, hilar lymphadenopathy, microcalcification, mediastinal mass, emphysema and pulmonary fibrosis — close to every change a chest X-ray can carry. Validation at scale came through the UK NHS SOM (System One Million) program.

**Annalise CTB** does the same for head CT, about 130 findings. ICH and LVO, plus microischemia, white matter disease, atrophy, aneurysm and venous sinus thrombosis.

Annalise's edge is **breadth of findings**. If Aidoc concentrates on urgent triage, Annalise is the "do not miss anything on this one image" safety net for routine reading. Per-case pricing keeps the marginal cost per patient low.

Chapter 6 · Qure.ai — Indian-born global TB and stroke standard

Mumbai-based **Qure.ai** was founded in 2016. Prashant Warier and Pooja Rao went after India and emerging markets from day one — tuberculosis, trauma, intracerebral hemorrhage. These are the most acute imaging problems in regions short on radiologists.

**qXR** — chest X-ray. Thirty-plus findings. After the WHO added AI-assisted chest X-ray to its active case-finding guideline for tuberculosis in 2021, qXR became the de facto global standard for that use case. It rides along with portable digital X-rays in TB screening camps in Vietnam, Nigeria and Indonesia. Funding flows through The Global Fund, Stop TB Partnership and USAID.

**qER** — head CT. ICH, fracture, midline shift, hydrocephalus, mass effect. Approvals across 40+ countries.

**qCT** — chest CT. Pulmonary nodule, emphysema, fibrosis.

Qure.ai's global signature is **infrastructure friendliness for low and middle-income countries**. Edge inference is tuned to run with unreliable internet in TB camps, and it integrates tightly with portable X-ray makers (Delft Imaging, GE Lumify). The result is that hospitals in Korean and Japanese cities and rural Indian TB camps run the same AI.

Chapter 7 · Viz.ai — redefining the clock on LVO triage

California-based **Viz.ai** was founded in 2016 by Chris Mansi (a UK neurosurgeon) and David Golan. The first product, **Viz LVO**, auto-detects large-vessel occlusion on head CT angiography (CTA) and pushes an alert to the neurointerventional team's phones. Its De Novo clearance in 2018 made it the first FDA-authorized device where AI directly notifies a clinician.

For LVO patients, time is brain. Standard workflow — CT, then radiology read, then neurology call, then mobilize interventional — averages 30 to 60 minutes. Viz LVO compresses that chain so that images from the gantry land in the interventionalist's hand within five minutes. Medicare recognized that time saved and applied NTAP in 2020.

The 2026 Viz.ai portfolio is broader.

- **Viz LVO / Viz CTP / Viz ICH** — stroke

- **Viz ANEURYSM** — aneurysm

- **Viz PE** — pulmonary embolism

- **Viz HCM** — hypertrophic cardiomyopathy

- **Viz AAA** — abdominal aortic aneurysm

- **Viz Subdural** — chronic subdural hematoma

Viz's second weapon is **Viz Connect** — a mobile collaboration layer that consolidates one patient's imaging and clinical data for the multidisciplinary team to share. A physician away from the hospital can review the non-contrast CT, CTA, CTP and lab results. AI alerts are button one; multidisciplinary collaboration is button two.

Chapter 8 · Rad AI · HeartFlow · Tempus · Arterys — workflow plus heart plus multi-omics

**Rad AI** — US. LLM-driven automation of the radiology reading workflow. Voice recognition plus GPT-4 generated report drafts shrink reading time. Rad AI Continuum launched in 2024. A regular keynote at RSNA and ACR.

**HeartFlow** — US. Simulates **FFRct (fractional flow reserve)** on coronary CT. It evaluates the hemodynamic significance of a stenosis non-invasively, sparing many invasive catheter studies. NICE in the UK NHS and Medicare both have reimbursement codes. IPO chatter through 2024.

**Cleerly Health** — US. Quantifies plaque characteristics (calcified, non-calcified, low-attenuation) on coronary CT. Built clinical evidence in collaboration with the ISCHEMIA and MESA cohorts.

**Tempus (Tempus AI)** — Chicago, US. Started in oncology as a **NGS sequencing + clinical data** integrator. IPO in 2024. Acquired **Arterys** (cloud cardiac MRI AI) in 2023 to bring imaging into the multi-omic stack. The Tempus vision is one patient's tumor NGS, imaging and drug response in the same chart.

**Caption Health** — acquired by GE HealthCare in 2023. AI-guided cardiac ultrasound acquisition. It lets a general nurse capture standard views.

**Ultromics** — Oxford, UK. Echocardiography for HFpEF risk prediction. EchoGo product line.

**Subtle Medical** — US. Improves MRI and PET image resolution and SNR with AI so scan times are halved. Reduces gadolinium contrast dose.

**Imagen Technologies** — US. X-ray in primary care, with AI plus remote radiology delivering an interpretation within 30 minutes.

Chapter 9 · Pathology AI — Paige · PathAI · Ibex · Owkin

Pathology digitized one step behind radiology. **WSI (Whole Slide Imaging)** scanners — Leica Aperio, Philips IntelliSite, Hamamatsu NanoZoomer — spread, more pathology departments accepted digital workflow, and AI moved in.

**Paige.ai** — New York, US. A spinout from Memorial Sloan Kettering (MSKCC). In 2021 **Paige Prostate Detect** became the first FDA-cleared prostate cancer AI. In 2024 the platform expanded to multi-cancer detection with **Paige PanCancer Detect**. Active deployment is in large hospitals and institutions with digital pathology workflows.

**PathAI** — Boston, US. Started in pharma partnerships — standardizing pathology evaluation in clinical trials — and expanded into the diagnostic workflow platform **AISight**. Became the standard in NASH (non-alcoholic steatohepatitis) trial scoring; IPO talk through 2024.

**Ibex Medical Analytics** — Israel. Focused on GI and breast biopsies. Galen Prostate, Galen Breast, Galen Gastric. Adopted by large European pathology departments.

**Owkin** — Paris, France. Federated-learning pathology and drug discovery. Patient data never leaves the hospital; only the model is shared across rounds. Partnerships with Sanofi and Bristol Myers Squibb.

Chapter 10 · Ophthalmology AI — Mediwhale · Verily · Eyenuk

The retina offers more clues to systemic disease per photograph than almost any other modality. In 2024 *Nature Medicine* and *The Lancet* published studies showing that retinal images can predict cardiovascular risk, diabetes status, and even Alzheimer's risk.

**Mediwhale** — Seoul, Korea. Founded in 2016. The flagship **Reti-CVD** predicts **cardiovascular risk (coronary artery calcium score)** from a single retinal photo. Holds Korean MFDS and EU CE approvals. In 2024 it was named in NICE Early Value Assessment candidates in the UK NHS.

**Verily** — an Alphabet subsidiary. Automated diabetic retinopathy (DR) screening — Automated Retinal Disease Assessment (ARDA). Deployed in primary care clinics in India and Thailand.

**Eyenuk** — US. **EyeArt** for automated DR screening. FDA cleared. Camera plus AI in primary care triages who needs an ophthalmology referral.

**IDx-DR** (Digital Diagnostics) — the first FDA PMA autonomous-diagnosis AI in 2018. Returns a "refer / not refer" decision without physician interpretation.

**Optos** — ultra-widefield retinal imaging plus AI.

Chapter 11 · Dermatology · endoscopy · surgical AI — SkinVision · AI Medical · Theator

**SkinVision** — Netherlands. A consumer smartphone app that evaluates melanoma suspicion from a photograph. CE Mark.

**DermaSensor** — US. Spectroscopy plus AI for non-invasive skin cancer assessment. Cleared by the FDA in 2024.

**Skinive** — Poland and Belarus. Primary-care skin cancer screening.

**AI Medical Service (AIM)** — Tokyo, Japan. Gastric and colonoscopy AI. One of the global leaders in endoscopy AI for gastric cancer. **gastroAI Model G** holds MHLW and PMDA approval and is expanding into Korea, Southeast Asia and Europe.

**Olympus EndoBRAIN** — Japan. Olympus's in-house colon polyp AI, naturally integrated with the endoscope hardware.

**Theator** — Israel. Automatically analyzes surgical video and indexes "what action happened when." Used in surgical education and quality assessment.

**C-SATS (Johnson and Johnson)** — surgical analytics. JOMICS / CSATS evaluate surgeon technique.

**Activ Surgical** — US. AR plus AI to synthesize fluorescence imaging intra-operatively for real-time visualization.

Chapter 12 · Genomic and multi-omic AI — Tempus · Foundation · GRAIL

**Tempus** — covered in Chapter 8. Imaging plus genomic plus clinical, integrated.

**Foundation Medicine** — a Roche subsidiary. **FoundationOne CDx** — NGS companion diagnostic for solid tumors. Wide FDA and MSAC approvals. Guides targeted therapy selection through variant detection.

**GRAIL Galleri** — a spinout from Illumina and then a separated entity. **Galleri** detects **more than 50 cancers early** from a single blood draw. The UK NHS-Galleri randomized trial (around 140,000 participants) is heading toward a 2026 readout that may be the single biggest inflection point in this category.

**Exact Sciences** — Cologuard (stool DNA for colorectal cancer), Oncotype DX.

**Guardant Health** — liquid biopsy ctDNA NGS.

The crux of multi-omic AI is putting one patient's imaging, genome, clinical notes, lab results and drug history into the same model and producing the next clinical decision. Data integration and privacy (federated learning) are the technical bottlenecks; FDA and CMS multi-omic companion diagnostic guidance is being clarified through 2025 and 2026.

Chapter 13 · Korean medical AI — Lunit · VUNO · JLK · Coreline · Deepnoid · Medical IP

Korea is one of the most active medical AI ecosystems globally. The MFDS published a separate AI medical device approval guide in 2017 and HIRA later defined reimbursement guidance, which gave the market a path.

**Lunit** — Chapter 4. KOSDAQ listed. Chest, breast, pathology.

**VUNO** — Seoul. Chest X-ray (VUNO Med-Chest X-ray), chest CT, bone age, brain MRI, ECG (VUNO Med-DeepECG). Listed on KOSDAQ in 2021. **VUNO Med-DeepCARS** is an inpatient cardiac arrest prediction system in use at Seoul National University Bundang Hospital and others.

**JLK Inspection** — focused on stroke. JLK-CTL (LVO), JLK-NCCT (ICH), JLK-CTP (perfusion), JLK-MRA. KOSDAQ listed in 2019.

**Coreline Soft** — chest CT. **AVIEW LCS Plus** is used in lung-cancer LDCT screening programs in Korea, the US and Europe. KOSDAQ listed in 2023.

**Deepnoid** — brain and spine imaging. Deep:NeuRAD, Deep:SPINE.

**Medical IP** — medical imaging simulation and 3D modeling.

**Standigm**, **Syntekabio** — drug discovery AI.

**Promedius** — bone density X-ray AI.

The MFDS reports around 250 AI medical device approvals in Korea as of 2024 statistics. Chest, neurology and musculoskeletal lead, with ophthalmology, cardiology and pathology catching up.

Chapter 14 · Japanese medical AI — AI Medical Service · LPixel · PFN Med · CureApp · Ubie

Japan is the home of gastric and colorectal endoscopy. The spring 2026 lineup of Japanese medical AI.

**AI Medical Service (AIM)** — Tokyo. Endoscopy global leader; see Chapter 11.

**LPixel** — a University of Tokyo spinout. EIRL aneurysm (brain aneurysm), EIRL Chest, EIRL Bone. Multiple PMDA approvals.

**PFN Med (Preferred Networks)** — Tokyo. The Chainer-era PFN moved into medical imaging and pathology.

**CureApp** — Tokyo. Digital therapeutics. Nicotine dependence app, hypertension app — PMDA-cleared and reimbursed under Japan's national insurance.

**Ubie** — Tokyo. AI intake plus a patient-facing symptom checker. In about 1,000 Japanese hospitals (free for hospitals) and consumer-facing with Ubie AI 症状検索.

**Caregia (M3 / MICIN line)** — telemedicine plus AI.

PMDA medical device review averages 11 to 15 months. A priority review track for SaMD (Software as a Medical Device) has been operating since 2023.

Chapter 15 · Regulation — FDA · MFDS · PMDA · EU MDR plus AI Act

A short paragraph on each regulatory frame.

**FDA (US)** — 510(k) for most clearances, De Novo for new categories, PMA for autonomous diagnosis. The 2024 finalization of the PCCP (Predetermined Change Control Plan) guidance opened the door to incremental retraining within a pre-agreed envelope.

**MFDS (Korea)** — first AI medical device guide in 2017. The 2023 revision tightened SaMD classification and change-management. Reimbursement runs through HIRA's new medical technology assessment (nHTA).

**PMDA (Japan)** — under MHLW. SaMD priority review. CureApp's reimbursed digital therapeutics attract global attention.

**EU MDR (EU)** — in force since May 2021. Applies to all medical devices. AI medical devices are mostly Class IIa/IIb and require Notified Body assessment.

**EU AI Act** — entered into force in 2024. Medical AI is almost universally classified as a "high-risk" system. MDR and the AI Act apply in parallel. Conformity assessment, data governance, human oversight, robustness and cybersecurity requirements.

**WHO** — guidelines for AI-based health tools. Standard operating procedures for AI chest X-ray in tuberculosis, malaria and other programs.

Chapter 16 · Data standards — DICOM · HL7 FHIR · OpenEHR · SNOMED CT

Medical AI's real weapon is data, not algorithms. Data only flows on top of standards.

**DICOM (Digital Imaging and Communications in Medicine)** — started as the ACR-NEMA standard in 1983. The storage and transmission format for almost all medical imaging. The standard path is for AI output to flow back into PACS as DICOM SR (Structured Report) or DICOM SEG (Segmentation).

**HL7 FHIR (Fast Healthcare Interoperability Resources)** — a REST + JSON standard for electronic health record exchange. R4 is the stable version. Epic, Cerner and Allscripts all support SMART on FHIR. AI output is written into the patient chart as a FHIR Observation resource.

**OpenEHR** — clinical data modeling standard. Used in Europe, Australia and parts of Korea.

**SNOMED CT** — clinical terminology system. Over 300,000 concepts. Codes diagnoses and findings.

**LOINC** — laboratory test codes.

**ICD-10 / ICD-11** — diagnosis classification.

Joining DICOM and FHIR is the medical AI day job. PACS (Sectra, Philips, GE, Fuji) goes to imaging AI goes to FHIR Observation goes to EHR (Epic, Cerner) goes to clinical workflow.

Chapter 17 · Hospital integration — Epic · Oracle Cerner · PACS vendors

**Epic Systems** — Wisconsin. About 40% of the US EHR market. Released GPT-4-powered clinical note summarization in 2024. Imaging AI integration runs through **Epic Cosmos** (federated data platform) and **SMART on FHIR**.

**Oracle Health (formerly Cerner)** — Oracle acquired Cerner for around USD 28.3 billion in 2022. Epic's strongest competitor. AI runs through Oracle's cloud and its own LLM linkages.

**Allscripts (Veradigm)** — outpatient EHR.

**PACS vendors**:

- **Sectra** — Sweden. Frequently ranked first in US radiology surveys.

- **Philips IntelliSpace** — Netherlands.

- **GE Centricity / Edison** — US. Expanding into an AI marketplace.

- **Fujifilm Synapse** — Japan.

- **Agfa Enterprise Imaging** — Belgium.

**Radiology worklist orchestration** — **Backbone AI / Rad AI / Aidoc aiOS** — manage multiple AI algorithms through a single worklist and prioritize the order in which one radiologist sees patients. A new market category in 2026.

Chapter 18 · Clinical LLMs — Med-PaLM 2 · GPT-4 plus DAX Copilot · Hippocratic · OpenEvidence

If imaging is the territory of computer vision, clinical notes, patient conversations and literature search are the territory of LLMs. The spring 2026 clinical LLM landscape.

**Med-PaLM 2 (Google DeepMind)** — medical Q&A. At the 2023 announcement, USMLE accuracy was reported around 86%, past the physician pass mark. Commercialized on Google Cloud as **MedLM**. Pilots at the Mayo Clinic and others.

**GPT-4 plus Microsoft Nuance DAX Copilot** — clinical speech recognition plus GPT-4 to auto-draft clinical notes. A physician-patient conversation converts directly into a SOAP note. Epic and Cerner integration. Used by roughly 20,000 US clinicians.

**Hippocratic AI** — US. A safety-first medical LLM. Slogan: "Gen-AI healthcare workforce." Replaces GPT-style models in patient monitoring, chronic disease coaching and medication counseling.

**Glass Health** — US. Clinical reasoning LLM. Auto-generates differential diagnoses.

**OpenEvidence** — US. Evidence-based search and answers grounded in medical literature. A partner of NEJM AI. Cuts a clinician's PubMed search time per question down to the minute scale.

**Abridge** — US. Clinical speech to auto-note. A competitor to DAX. Official Epic partner.

**Nabla Copilot** — France. A European clinical voice LLM.

Chapter 19 · Federated learning plus privacy — Owkin · NVIDIA FLARE · Rhino Health

Medical AI's data problem is simple. One hospital's data is not enough for generalization, and pooling data across hospitals brings the patient inside the dataset. The answer is **federated learning** — data never leaves the hospital; only model updates are exchanged each round.

**Owkin** — see Chapter 9. The European leader in federated learning for pathology and drug discovery.

**NVIDIA FLARE (Federated Learning Application Runtime Environment)** — NVIDIA's open-source federated learning framework. Paired with MONAI, it is the de facto standard for medical imaging.

**Rhino Health (US)** — sells the federated learning platform itself as SaaS. Imaging AI developers subscribe to gain data access.

**MELLODDY (EU / IMI consortium)** — about a dozen pharma companies train models on their compound libraries without ever sharing the compounds.

The stack mixes **DP (Differential Privacy)**, **HE (Homomorphic Encryption)**, **SMPC (Secure Multi-Party Computation)**, **TEE (Trusted Execution Environment, SGX, SEV, H100 Confidential)** and federated learning itself. The trade-off between mode and intensity is data sensitivity against compute cost.

Chapter 20 · The clinical workflow in one page — five minutes inside an ED head CT

Back to the opening scenario. Walk the five minutes of one ED head CT case minute by minute.

- **T+0:00** — patient arrives, NIHSS 4, ED physician orders non-contrast head CT and CTA.

- **T+1:30** — patient off the gantry. The DICOM series flows to PACS (Sectra).

- **T+1:45** — the PACS DICOM router fans the study out to Aidoc aiOS and the Viz.ai server simultaneously.

- **T+2:30** — Aidoc BriefCase ICH evaluates the non-contrast study for intracranial hemorrhage. Result: negative. Standard worklist priority.

- **T+2:45** — Viz LVO scores the CTA for a right MCA M1 occlusion. Result: positive. Viz Connect pushes the alert to the neurointerventional team and on-call neurology.

- **T+3:30** — the radiology resident clicks the red-banner case in the worklist. Reviews CT and CTA together.

- **T+4:00** — the neurointerventional physician opens Viz Connect on their phone. Calls the interventional suite to prep.

- **T+4:30** — the resident dictates a read. Microsoft DAX converts the dictation into a SOAP note and writes it to the Epic chart.

- **T+5:00** — the patient is already on the way to the interventional suite.

Every arrow in that flow has a company behind it. PACS (Sectra), triage (Aidoc), LVO (Viz.ai), multidisciplinary collaboration (Viz Connect), voice-to-note (DAX), EHR (Epic). One patient's five minutes traverses the entire medical AI ecosystem.

Chapter 21 · How to start — by role: radiologist, clinician, developer, administrator

The entry point into this field depends on your role.

**Radiologist / clinician** — start by comparing AI outputs from tools already installed on your PACS (Aidoc, Lunit, Viz) against your own reads. The fastest learning loop you can run without IRB or consent overhead. Follow the AI tracks at RSNA, ECR and KCR each year.

**Developer / data scientist** — start with **MONAI** (NVIDIA's medical imaging PyTorch), **fastMONAI**, **TorchIO** and **nnU-Net**. Train your first model on public datasets — **NIH ChestX-ray14**, **MIMIC-CXR** (MIT), **CheXpert** (Stanford), **TCIA** (Cancer Imaging Archive), **MedMNIST**, **BraTS** and **LIDC-IDRI**. The Kaggle RSNA challenge (annual since 2017) poses a different imaging problem every year.

**Administrator / hospital ops** — the first question on imaging AI procurement is "does it integrate with my PACS?" The second is "which workflow KPI (TAT, triage accuracy, report turnaround) does it move?" The third is "is there reimbursement, NTAP or HIRA coverage?" Without a clinical champion (one radiology head), adoption does not stick.

**Regulatory / policy** — track the FDA's [AI/ML-Enabled Medical Devices list](https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices), Korea's MFDS [AI medical device library](https://www.mfds.go.kr/) and PMDA's SaMD page on a regular cadence.

**Patient** — your right to know whether your study went through an AI assist is getting clearer. The EU AI Act places a disclosure obligation on the healthcare provider.

Chapter 22 · Limits — data bias, generalization, accountability

The last chapter is an honest word about limits.

**Data bias** — almost every medical AI model was trained on large academic-hospital data in the US, Europe and East Asia. Distribution bias exists by race, sex, age and equipment vendor. Reports of chest X-ray AI underperforming in Black patients (JAMA Network Open 2021) and mammography AI underperforming in dense breast tissue have stacked up.

**Generalization limits** — performance degrades on images from a vendor, resolution or protocol the model has not seen. Domain adaptation, test-time adaptation and synthetic data via MONAI Generative help but there is no general solution.

**Accountability** — when an AI influences a decision and the outcome is poor, who is responsible? The general position of the FDA, MFDS and PMDA is that AI is a clinician aid; the final decision rests with the human. But in time-critical use cases like LVO triage, where alerts go before a human reads, the boundary blurs.

**Black box** — interpretation of imaging AI decisions still relies on proxy explanations — Grad-CAM, attention maps, SHAP. A causal answer to "why did the model call this positive?" remains hard for the clinician.

**Friction between regulation and update** — until PCCP settles in, once a model was cleared the cleared weights had to stay frozen. Data flows, clinical practice shifts, yet the model stood still. That contradiction is being resolved through 2026.

These limits are not reasons to dismiss the field. The same kinds of limits existed in the 1990s digital imaging transition and the 2000s PACS rollout, and medicine absorbed them while moving forward. AI follows the same arc — critically, and step by step.

Chapter 23 · Conclusion — imaging plus genome plus clinical, in one chart, for the next ten years

Spring 2026, five minutes inside one ED head CT case. We saw a picture of the era. PACS, imaging AI, LVO triage, multidisciplinary collaboration, voice LLM, EHR — different companies, different standards, different algorithms — converging on one point, one patient's chart.

The next ten years point clearly. **Imaging, genomics, clinical notes, labs, drug history and wearables joined on the same patient chart, along the same timeline** — multi-omic AI. Tempus and Foundation Medicine, Owkin and Lunit SCOPE are the first beacons. On top of that, **living models under PCCP**, **federated learning that respects privacy**, and **a global regulatory harmonization where the EU AI Act and FDA PCCP cooperate**.

The radiologist does not disappear. If anything, one physician handles more patients, more modalities, and influences more decision points. AI is a tool that extends a clinician's cognitive bandwidth, the way glasses converted a person at 0.1 visual acuity to 1.0. Five minutes per patient does not become five seconds. Five minutes lets a physician know far more.

And inside that arc you find Lunit in Korea, AI Medical in Japan, Qure in India, Aidoc in Israel, Annalise in Australia, and Viz, Tempus and Paige in the US together. Medical AI is not a one-country game.

Chapter 24 · References

- [FDA · Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices](https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices)

- [FDA · Predetermined Change Control Plans for AI/ML Medical Devices Guidance (2024)](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence)

- [Aidoc · Official Site](https://www.aidoc.com/)

- [Lunit · INSIGHT CXR / MMG / DBT / SCOPE](https://www.lunit.io/)

- [Annalise.ai · CXR / CTB](https://annalise.ai/)

- [Qure.ai · qXR / qER / qCT](https://www.qure.ai/)

- [Viz.ai · Stroke and Beyond](https://www.viz.ai/)

- [Tempus AI · Multi-Modal Oncology](https://www.tempus.com/)

- [HeartFlow · FFRct](https://www.heartflow.com/)

- [Paige.ai · FDA-cleared Prostate AI](https://www.paige.ai/)

- [PathAI · AISight Platform](https://www.pathai.com/)

- [Mediwhale · Reti-CVD](https://www.mediwhale.com/)

- [Verily · Diabetic Retinopathy Screening (ARDA)](https://verily.com/)

- [AI Medical Service · Endoscopy AI](https://www.ai-ms.com/)

- [Hippocratic AI · Healthcare LLM](https://www.hippocraticai.com/)

- [Med-PaLM and MedLM · Google Research](https://sites.research.google/med-palm/)

- [Microsoft DAX Copilot for Clinicians](https://www.microsoft.com/en-us/health-solutions/clinical-workflow/dragon-copilot)

- [MONAI · Medical Open Network for AI (NVIDIA)](https://monai.io/)

- [HL7 FHIR · Specification](https://hl7.org/fhir/)

- [DICOM Standard](https://www.dicomstandard.org/)

- [EU AI Act · Official](https://artificialintelligenceact.eu/)

- [WHO · Use of AI for Tuberculosis Screening (chest X-ray)](https://www.who.int/publications/i/item/9789240029033)

- [Korean MFDS · AI Medical Device Library](https://www.mfds.go.kr/)

- [Japanese PMDA · Medical Device Review](https://www.pmda.go.jp/english/)

- [The Lancet Digital Health · Medical AI validation papers](https://www.thelancet.com/journals/landig/home)

- [RSNA · Radiological Society of North America (annual AI track)](https://www.rsna.org/)

- [KCR · Korean Congress of Radiology](https://www.radiology.or.kr/)