Inside the hospital, the range of things AI can touch is growing. It can reconstruct the pixels of a CT scan, identify anatomy in surgical video, design a candidate molecule, carry supplies along a ward and, perhaps one day, predict how a cell will react to treatment. A cluster of announcements in Tokyo in July 2026 presented these not as isolated demonstrations but as parts of a connected industrial stack linking computing, robots, images, molecules and biological data.
The immediate catalyst was NVIDIA’s July 15 account of projects with Japan’s healthcare and life-sciences ecosystem. Kawasaki Heavy Industries plans to use NVIDIA platforms in surgical support, nursing assistance and hospital transport robots. Direava is developing a surgical vision-language model, or Surgical VLM, intended to interpret video in real time and answer a clinician’s questions in natural language. Canon Medical Systems has launched Japan’s first domestically produced photon-counting CT system, while NVIDIA says Fujifilm has commercialized the country’s first whole-body CT powered by Blackwell GPUs and diffusion-based reconstruction.
Beyond the hospital, Eisai has joined Tokyo-1, the pharmaceutical AI supercomputer operated by Xeureka, alongside founding members Astellas Pharma, Daiichi Sankyo and Ono Pharmaceutical. SyntheticGestalt has released a molecular representation model called ZAO and a ligand-generation model called KOYA. Takeda and Boltz are moving generative drug-design models toward industrial use. Biomy, meanwhile, is building what it calls a Virtual Cell: a foundation model intended to learn the tumor microenvironment from clinical tissue data supplied through the Japanese Foundation for Cancer Research.
First, separate four levels of maturity
The phrase healthcare AI obscures more than it reveals unless the underlying projects are divided into at least four groups. The first consists of imaging systems already on the market. The second includes medical devices and software cleared for tightly defined supportive uses. The third covers systems under evaluation in a hospital or surgical environment. The fourth consists of experimental drug-discovery and cell models used to generate research hypotheses.
| Stage | Examples in 2026 | What they can do—and what remains unproven |
|---|---|---|
| Commercial systems | Canon Ultimion; Fujifilm’s AI-reconstruction CT | Acquire and reconstruct images. Diagnostic performance and patient benefit must still be established for each use and protocol. |
| Approved assistance | Hinotori; Direava’s Kinosura; the EndoBRAIN family | Assist a trained clinician. Approval does not imply autonomous surgery or diagnosis. |
| Clinical or operational evaluation | FORRO, Nurabot, Surgical VLM, next-generation surgical support | Being tested in real hospital environments. A single-site result cannot establish general performance. |
| Research infrastructure | Tokyo-1, ZAO, KOYA, Boltz models, Virtual Cell | Prioritize molecules, targets or experiments. They do not establish efficacy or safety in people. |
This is not merely cautious labeling. In medicine, a model’s benchmark accuracy and its effect on complications, staff workload or quality of life are different questions. Laboratory performance, regulatory authorization, reimbursement, hospital procurement and routine clinical benefit are separate gates. A technology has to pass each one.
The history of surgical robots: extending the hand before automating it
Robot-assisted surgery did not begin with a machine deciding how to operate. It spread as a form of telemanipulation: the surgeon’s hand movements are translated into small instruments, tremor can be filtered, and a three-dimensional view helps the operator work in a confined space. Japan’s public insurance system began covering robot-assisted prostate surgery in 2012 and extended coverage into gastrointestinal surgery and other fields in 2018.
A domestic milestone came in 2013, when Kawasaki Heavy Industries and Sysmex established Medicaroid in Kobe. The venture combined Kawasaki’s experience in mechanisms and control with Sysmex’s medical-device expertise. Its hinotori Surgical Robot System received Japanese approval in August 2020 and entered clinical use that year. Its indications expanded into general surgery and gynecology in 2022 and thoracic surgery in 2024; by 2026, Medicaroid had taken steps into Europe.
Yet hinotori, like the widely used da Vinci system, remains fundamentally surgeon-controlled. The robot moves instruments, but it does not independently decide what tissue to cut and complete an operation. Kawasaki’s new plan to use NVIDIA Holoscan IGX, Isaac for Healthcare, Isaac GR00T and Cosmos is framed as development of surgical support. It is not an announcement that autonomous surgery has entered Japanese clinical care.
On July 16, Kawasaki and Fujitsu also said they would study a hospital solution connecting electronic medical records, robot operations and AI. The hardest part may not be the robotic arm. It may be ensuring that surgical schedules, tests, supplies, beds and staff permissions are joined safely, without letting an AI retrieve the wrong patient, order or medication.
Turning surgical video into language
Direava’s Surgical VLM brings a vision-language model into the operating room. The concept is to interpret anatomy and the state of the surgical field from endoscopic video, then respond conversationally when a clinician asks what structure is visible or what deserves attention. The same architecture could eventually support navigation, training and documentation.
Direava says it conducted a live-environment evaluation during gastric-cancer surgery at Keio University Hospital on February 20, 2026. The company reported 84.7% anatomical accuracy, 82.9% clinical usefulness and 97.4% linguistic fluency. Those are promising early measures, but they came from a company-defined, single-site evaluation. They are not evidence that the system reduces complications or operating time. External testing must show whether performance holds across hospitals, procedures, cameras, smoke, bleeding and varied patient anatomy.
The company already has a narrower precedent. Kinosura, software that recognizes and highlights the recurrent laryngeal nerve in surgical video, received Japanese medical-device approval in December 2025 for robot-assisted surgery for malignant esophageal tumors. A focused assistive model has a defined input, output and user. A general model that answers broad questions creates harder problems: when it should abstain, how it cites visual evidence, how updates are controlled and how a clinician can recognize a plausible but false answer.
Hospital robots may create value by carrying, not cutting
The near-term impact of medical robotics may appear first in a cart rather than a scalpel. Kawasaki’s FORRO delivery robot is intended to move drugs, specimens and supplies. The company conducted preparatory work at Tokyo Medical University Hospital from fiscal 2023 and began a new demonstration in February 2025. The robot is designed to navigate hallways around people and obstacles while working with elevators and automatic doors.
Nurabot, jointly developed with Foxconn, supports tasks around nursing such as meal service, medicine and baggage delivery and visitor guidance. Trials began at Taichung Veterans General Hospital in Taiwan in April 2025, with market introduction targeted for fiscal 2026. Together with the humanoid Nyokkey concept, the practical objective is less to replace nurses than to reduce walking and repetition so staff can spend more time observing, explaining and caring for patients.
A movement success rate is not enough to establish hospital value. Infection control, cleaning, charging, nighttime noise, emergency passage, elevator delays, fall and collision risk, cybersecurity and access separation from the electronic record all matter. A robot operating successfully is not the same thing as a ward experiencing less work. The missing step is redesigning the workflow around it.
Two CT revolutions: detector physics and reconstruction AI
Medical imaging contains two distinct innovations that can easily collapse into a single “AI CT” headline. Canon Medical’s Ultimion, launched on April 17, 2026, is the first photon-counting CT developed and manufactured in Japan. Conventional CT detectors integrate X-ray energy into a combined signal. A photon-counting detector counts individual X-ray photons and can retain information about their energy. Depending on the protocol, that can improve spatial or spectral resolution and dose efficiency.
The central advance is detector physics, not AI alone. Canon acquired Canadian detector specialist Redlen Technologies in 2021, installed a research whole-body photon-counting system at the National Cancer Center Hospital East in 2022 and began clinical research in 2023 before commercial launch. GPU computing helps reconstruct the enormous data stream, but “photon counting” should not be treated as a synonym for artificial intelligence.
Fujifilm represents the other path. NVIDIA says the company commercialized Japan’s first whole-body CT powered by Blackwell and diffusion-based deep-learning reconstruction. A diffusion model iteratively recovers an image from noisy data, potentially supporting lower-dose or faster scans while preserving useful detail. But generative reconstruction must be tested carefully to show that it neither erases subtle lesions nor invents structures, across devices, body regions and acquisition settings.
An image that looks cleaner is not automatically a more accurate diagnosis. Comparative studies need to measure lesion detection, false positives, repeat scans, reading time, radiation dose and patient outcomes separately.
Tokyo-1: rival drugmakers share the computing layer
The other major arena for healthcare AI lies years before a medicine reaches a patient. Xeureka, established by Mitsui in 2021, put Tokyo-1 into full operation in February 2024. Built around NVIDIA DGX H100 infrastructure, the project began with Astellas Pharma, Daiichi Sankyo and Ono Pharmaceutical as pharmaceutical partners sharing computing resources and technical work.
Eisai joined in April 2026. A July announcement said the group had held more than 60 working-group meetings and completed five technical validation themes. The companies are not pooling their most valuable drug candidates. They are jointly testing common methods and infrastructure for molecular simulation, generative AI and large-scale data processing—the foundational work that would otherwise be repeated behind each corporate wall.
Astellas has used DGX and BioNeMo resources for models including astABpLM, a language model intended to predict antibody properties. Daiichi Sankyo is conducting ultra-large virtual screening. Ono is using Boltz-2 to predict protein-ligand interactions, while Xeureka works on automated discovery workflows. The strategic value of Tokyo-1 is not simply shared GPU purchasing. It is the possibility of reducing duplicated infrastructure work and helping scientists reach experimentally testable hypotheses sooner.
AI that makes molecules: a faster proposal is not yet a drug
SyntheticGestalt’s ZAO is a foundation model designed to learn a molecular representation that includes multiple possible three-dimensional conformations—what the company calls a “4D” view. SyntheticGestalt says it pre-trained ZAO on roughly 10 billion compound records and achieved leading results on tested public benchmarks. KOYA generates candidate ligands while accepting medicinal-chemistry constraints such as synthesizability. Both can be called through NVIDIA’s BioNeMo Agent Toolkit.
BoltzMol-1 and BoltzProt-1, which Boltz is working to deploy with Takeda, apply generative modeling to small-molecule and protein design. Agentic systems can string together database search, structure prediction, generation, scoring and the selection of another computation. But a molecule rated highly by a model is not guaranteed to enter a cell, avoid toxicity, reach the right tissue or work in a human body.
The part AI can most readily compress is the computational narrowing of a vast search space. Synthesis, laboratory validation, preclinical evaluation, manufacturing, clinical trials and regulatory review remain physical bottlenecks. A more meaningful metric than benchmark rank is the share of AI proposals that can be synthesized, reproduce experimentally and reduce the number of design-test cycles compared with the existing process.
What is a Virtual Cell? Not a patient’s digital clone
The most futuristic phrase in the 2026 announcements is biomy’s Virtual Cell. The company plans to combine gene expression, protein expression and pathology images from clinical tissue supplied through the Japanese Foundation for Cancer Research, learning how tumor, immune and stromal cells relate to one another in space. A partnership with 10x Genomics adds spatial-transcriptomics datasets generated with the Xenium platform.
A virtual cell is not a digital person that reproduces a particular patient’s whole body. It is a computational model of observed cell and tissue states. Its purpose is to predict how gene expression or cell populations might change after a gene is suppressed, a drug is introduced or an immune environment shifts. If successful, it could prioritize hypotheses about target validity, treatment response and resistance before researchers commit to expensive experiments.
Biomy reported that NVIDIA single-cell RAPIDS reduced processing time by as much as 90% in one internal spatial-multiomics analysis workflow compared with CPU processing. That does not mean drug development became 90% faster. It describes acceleration of a particular processing step under particular data, hardware and test conditions. Computational speed and biological validity are independent measures.
The deeper challenge is that causality is buried inside biology. A tissue sample changes with sampling location, treatment history, age and preservation. Spatial datasets are rich but expensive and incomplete; researchers cannot observe every cell in a tumor at every point in time. A model necessarily infers what was not observed. Its predictions therefore have to remain falsifiable through new tissue samples, perturbation experiments, preclinical studies and, ultimately, clinical trials.
From 2003 to 2026: genome, iPS, single cells, virtual cells
The virtual-cell idea did not appear from nowhere. In 2003, the international Human Genome Project, with participation from Japan, completed the reference sequence of roughly three billion DNA bases. That sequence became a common map. It described much of the parts list, but not which genes operate in which cell, at what time and under what environmental conditions.
In 2006, Shinya Yamanaka and colleagues at Kyoto University reprogrammed mature mouse cells into induced pluripotent stem cells. They reported human iPS cells in 2007, and Yamanaka shared the Nobel Prize in Physiology or Medicine in 2012. The ability to build disease-relevant cells from patient tissue opened a path for testing drugs in a dish and for moving repeatedly between experimental and computational models.
During the 2010s, single-cell RNA sequencing separated cell populations that had previously been averaged together. Spatial transcriptomics added not only which genes were active but where they were active within a tissue. In 2026, virtual-cell projects are attempting to join the genome’s reference map, iPS experimental systems, single-cell observations and spatial data with foundation models.
| Year | Milestone | Why it matters now |
|---|---|---|
| 2003 | Human Genome Project completed | A shared reference sequence for biomedical research. |
| 2006–07 | Mouse, then human iPS cells | Patient-derived disease models and a new route into drug research. |
| 2012 | Japanese insurance coverage begins for robot-assisted surgery | Surgical robotics moves from research toward routine care. |
| 2014 | Japan regulates diagnostic and treatment software as medical devices | A legal route for reviewing clinical AI. |
| 2018 | EndoBRAIN approved | Japan’s first approved AI-assisted endoscopy software. |
| 2020 | Hinotori approved; DASH for SaMD begins | A domestic surgical robot and software-review reform advance together. |
| 2024 | Tokyo-1 enters full operation | A shared pharmaceutical AI computing platform. |
| 2026 | Surgical VLM, next-generation CT and Virtual Cell | Images, language, molecules, cells and robots begin to connect. |
Japan’s clinical AI history: the narrow, assistive path of EndoBRAIN
A key predecessor in Japan is EndoBRAIN, developed through a group that included Showa University, Nagoya University and Cybernet Systems. Approved in December 2018, it was trained on about 60,000 endoscopic images to help a physician judge whether a colorectal lesion was neoplastic. Olympus began sales in 2019. EndoBRAIN-EYE, trained on about 3.95 million images to assist lesion detection, gained approval in 2020.
The original EndoBRAIN did not continue learning and changing after sale. Regulators reviewed a fixed algorithm for a defined organ, image input, output and user. That narrow assistive design made performance and risk easier to measure. It established a pattern that continues in today’s approved surgical-video tools.
Foundation models are moving in the opposite direction: one model can answer many questions and update quickly. The broader the task, the harder it becomes to decide which change requires another review, how performance drift should be monitored at each hospital and how free-text errors should be recorded. Technology is becoming more general; medical safety still demands a precise intended use. That tension will define the next regulatory period.
Regulation, reimbursement and cybersecurity: approval is not the finish line
Japan’s Pharmaceuticals and Medical Devices Act, effective in its present form from November 2014, brought standalone diagnostic and treatment software into medical-device regulation. The Ministry of Health, Labour and Welfare launched DASH for SaMD in 2020 and strengthened consultation and review in 2021. The Pharmaceuticals and Medical Devices Agency began a priority-review trial for eligible software in 2022 and made a continuing scheme available from fiscal 2026.
A review examines intended use, training and evaluation data, performance, safety and the division of labor with the clinician. Adoption, however, also requires reimbursement, procurement, integration, staff training and maintenance. The EndoBRAIN experience has shown the commercial gap that can remain when it is unclear how an approved AI changes reimbursement or hospital work.
More connectivity also creates more cyber risk. Japan incorporated cybersecurity into the essential requirements for medical devices, with compliance required after the transition period in April 2024. As surgical robots, CT scanners, electronic records, clouds and transport robots approach the same network, updates, authentication, logs, software bills of materials, vulnerability response and a safe state during communications failure have to be designed across the hospital—not only inside each product.
Why Japan? Aging and workforce pressure create urgency
As of October 2024, Japan had 36.243 million people aged 65 or older, 29.3% of the population, according to the Statistics Bureau. The population over 75 will continue to grow, while the working-age population declines. The health ministry expects inpatient demand in some regions to peak around 2040 and home-care demand later still.
Japan also began physician overtime limits in April 2024. The general A-level cap is 960 hours of overtime and holiday work per year, with designated exceptions for regional care, training and other needs extending as high as 1,860 hours. AI and robots are attractive because demand and labor supply are moving in opposite directions.
AI does not automatically resolve a shortage. If it creates more false alarms, data entry, troubleshooting or patient explanations, work changes shape rather than disappearing. An implementation should measure staff walking, interruptions, documentation time, overtime and retention, together with patient waiting and safety—not simply the number of AI transactions.
The largest risks are invisible in an accuracy score
The first is dataset bias. A model trained on clean images from a large specialist hospital may not perform equally on older equipment or a different patient population in a regional hospital. Rare conditions, children, older adults, pregnant patients and people with multiple illnesses may be underrepresented.
The second is automation bias. The more fluent a model’s prose and the more precise its visual overlay, the easier it is to over-trust. A rule that the physician remains responsible is not enough. Systems need useful uncertainty, evidence displays, an easy way to dissent, training in error detection and studies that compare clinicians with and without the AI.
The third is privacy and data governance. Combining pathology, genomes and clinical records increases the possibility of re-identification. Patients need clarity on permitted uses, overseas cloud processing, retention, model memorization and whether a research withdrawal can be honored after training.
The fourth is platform dependence. Many of the 2026 projects connect to NVIDIA GPUs, CUDA, BioNeMo, Holoscan and Isaac. A common platform speeds development, but it also exposes users to price, supply, compatibility and export-control risks. Hospitals and research organizations need standards and contracts that let them move models and data to another environment.
Ten questions for 2026–2030
Healthcare progress cannot be counted in announcements or GPUs. Over the next four years, Japan’s programs should be judged by whether they can answer the following questions with public evidence.
| Area | Evidence that should be reported |
|---|---|
| Surgical support | External-site sensitivity and specificity, complications, operating time, incorrect guidance and clinician overrides. |
| Hospital robots | Not only delivery success, but staff walking, overtime, interruptions, collisions, downtime and total cost. |
| CT | Dose, lesion detection, false positives, repeat scans and interpretation time against conventional reconstruction. |
| Drug-discovery AI | Synthesis rate, experimental reproduction, design cycles, and time and cost from target to clinical testing. |
| Virtual Cell | Predictions on held-out patients and sites, falsification through perturbation, failed cases and calibrated uncertainty. |
| Equity | Performance by age, sex, region, device and disease group. |
| Human oversight | Accountability, stop authority, post-update evaluation and incident reporting. |
| Cybersecurity | Time to remediate vulnerabilities, safe offline behavior, access audits and recovery exercises. |
| Economics | Reimbursement after approval, total hospital ownership cost and availability outside major centers. |
| Patient benefit | Survival, complications, waiting, quality of life and confidence in explanations as final outcomes. |
Conclusion: Japan is not building a substitute for the physician
Japan’s healthcare AI push is not accelerating because of one invention. A domestic surgical robot approved in 2020, endoscopic AI dating to 2018, Tokyo-1 in 2024, and the CT, surgical VLM, hospital robot and Virtual Cell projects of 2026 are beginning to connect on common computing platforms. The AI that reads an image, the robot that moves an instrument and the model that designs a molecule are drawing closer.
The largest possibility is to turn medicine from a one-way sequence—find disease, then treat it—into a loop of observation, prediction, intervention and learning. Surgical video could return to training. Clinical tissue could improve a virtual cell. Experimental failures could refine the next molecular proposal. If that loop is made safe, scarce expertise could reach more hospitals and researchers.
Medicine, however, is judged not by the most persuasive demonstration but by its ability to protect the most vulnerable patient. A robot can extend a surgeon’s hand without assuming responsibility. A virtual cell can choose an experiment without treating a person. The true infrastructure is the system that forces AI-generated hypotheses through skepticism, regulation, trials and routine clinical evidence.
The 2026 announcements show that Japan is assembling the components of a healthcare AI stack. The next test is whether it can build the interoperability, institutions and evidence that connect those components to patient benefit. The final circuit between a surgical robot and a virtual cell is not a GPU. It is trust.
Sources and further reading
- NVIDIA: Japan’s AI Ecosystem Advances Healthcare and Life Sciences — the July 2026 overview of healthcare, pharmaceutical, robotics, imaging and Virtual Cell projects.
- Kawasaki Heavy Industries and Fujitsu: hospital solution business study — linking records, robotics and AI.
- Kawasaki Heavy Industries: FORRO hospital-delivery trial — operational testing at Tokyo Medical University Hospital.
- Kawasaki Heavy Industries: Nurabot nursing-assistance robot — development with Foxconn and Taiwan trials.
- Medicaroid: company history — founding in 2013 and hinotori approval, clinical use and indication expansion.
- Direava: live-environment Surgical VLM evaluation — the Keio University Hospital test and company-reported metrics.
- Canon Medical Systems: press releases — the April 2026 launch of the domestically produced Ultimion photon-counting CT.
- Canon: Japan’s first domestically produced photon-counting CT research system — Redlen detector technology and installation at the National Cancer Center Hospital East.
- Xeureka: Tokyo-1 full-scale operation — February 2024 launch, DGX H100 infrastructure and founding pharmaceutical participants.
- Xeureka: Eisai joins Tokyo-1 — April 2026 participation and working-group progress.
- NVIDIA and Astellas Pharma case study — antibody modeling, BioNeMo and Tokyo-1 use.
- Biomy: Virtual Cell development — clinical cancer data, spatial multiomics and the limits of its processing-speed claim.
- Biomy and 10x Genomics: Xenium collaboration — spatial-transcriptomics data for foundation models.
- SyntheticGestalt: ZAO foundation model — multi-conformation molecular representations and nvMolKit.
- AMED: EndoBRAIN medical-device approval — 2018 approval, training images and assistive intended use.
- AMED: EndoBRAIN-EYE approval — deep-learning detection support and training data.
- PMDA: Review of Software as a Medical Device — the PMD Act, SaMD review and priority pathways.
- Ministry of Health, Labour and Welfare: DASH for SaMD — Japan’s software medical-device acceleration policy.
- PMDA: Medical-device cybersecurity — essential requirements and the April 2024 compliance date.
- Statistics Bureau of Japan: Population Estimates 2024 — 36.243 million people aged 65 or older, 29.3% of the population.
- Ministry of Health, Labour and Welfare: physician work-style reform — overtime limits effective from April 2024.
- NHGRI: International consortium completes the Human Genome Project — the 2003 reference-sequence milestone.
- Kyoto University CiRA: Shinya Yamanaka awarded Nobel Prize — mouse iPS cells in 2006, human iPS cells in 2007 and their relevance to drug discovery.
Editor’s note: Product plans and performance figures are based on company statements available in July 2026. This report does not treat vendor benchmarks as evidence of improved patient outcomes and distinguishes approved uses, hospital evaluations and experimental research.