Today, after a patient completes a chest CT scan, clinicians pull up the images, review prior test results and treatment histories, and interpret what the new findings signify. Tomorrow, the next wave of trillion-dollar medical AI companies will likely emerge from three pillars: clinician workflow systems used day-to-day, datasets spanning the entire patient care journey, and foundational models that continuously deliver new diagnostic capabilities.
Tracing this trajectory reveals three global hubs. Pro Medicus (ASX: PME) in Melbourne, Australia, built its business around imaging workflows; Tempus AI (NASDAQ: TEM) in Chicago, US, focuses on diagnostics and patient data; Diagens Tech (HKEX: 2526) in Hangzhou, China, innovates with medical imaging foundational models. As of 8 October 2026, their respective market capitalisations stood at approximately RMB 78 billion, RMB 84 billion, and RMB 22 billion.
Each company began with a distinct business focus: Pro Medicus primarily sells imaging software; Tempus generates most of its revenue from molecular diagnostics and data services; Diagens Tech deploys medical imaging large foundational models into hospitals and clinical practice. Together, they illuminate a critical industry question: if workflows and data already command high valuations, could models that continuously generate clinical capabilities become the third high-value asset class in medical AI?
This article argues that Diagens Tech is positioning itself to occupy exactly this niche. By expanding from one regulated use case to additional specialties and hospital sites, it empowers clinical specialists to build, deploy and operate thousands of custom models on one unified foundation. As AI infinitely amplifies specialist expertise, medical imaging foundational models evolve from standalone technology products into scalable platform assets.
A Paradigm Shift: Health Demands Answers, and Above All, Evidence
The arrival of large models in healthcare reshapes how people engage with health information. In 2026, OpenAI rolled out health features that connect to medical records and Apple Health with user consent, reporting that more than 300 million people submit health queries to ChatGPT each week. Health management may shift from sporadic web searches to continuous insight, reminders and follow-up queries. Even so, while large language models can interpret reports, they cannot validate tumour shrinkage or organ damage from that data alone, let alone deliver reliable forecasts of disease progression.
On another frontier, AI is transforming drug discovery. Google DeepMind’s AlphaGenome examines how genetic variants alter molecular properties; NVIDIA and Eli Lilly announced up to USD 1 billion joint investment over five years; Rentosertib, a candidate drug co-developed by Insilico Medicine, entered Phase Ⅲ trials in 2026. From molecular hypotheses to clinical trials, models now contribute to discovery and R&D decisions. Health management seeks to understand what occurs within the body, while drug development explores how treatments alter disease outcomes. Both rely on reproducible, observable clinical evidence.
As the primary source of clinical evidence, medical imaging covers roughly 14,000 indication-specific tasks across CT, MR, ultrasound, endoscopy, pathology, microscopy, dermatoscopy and fundus imaging, accounting for 80% of all global clinical data. According to the World Health Organization, over 25 billion medical imaging scans are performed worldwide each year. This massive volume makes imaging one of the richest, most reliable observation layers within healthcare systems. Medical records, comparatively, are human-language documentation of those core observations.
Model development and clinical trials are converging in medical imaging. Google’s MedGemma 1.5 delivers interpretation for 3D CT, MRI and pathology images, while Microsoft’s CARE‑X focuses on chest X-ray reading and lesion localisation. The Swedish MASAI randomised trial enrolled approximately 106,000 women. Within a defined breast screening pathway, the AI-assisted workflow achieved higher breast cancer detection sensitivity (80.5% versus 73.8%), with comparable false positive control (98.5% specificity in both arms). Advances in technical performance and clinical evidence are converging. The next question defining industry value centres on integrating these capabilities into clinicians’ daily tools and building sustainable revenue streams.
Pro Medicus’ Screens in Melbourne: Imaging Information Management Software — Embedded in Clinicians’ Daily Workflows
The “screens” referenced here refer to end‑to‑end workflow management software for clinicians to view, compare and analyse medical images. Pro Medicus’s Visage7 integrates image viewing, archiving and workflow management. Clinicians open, review and compare imaging data on this system every day. Going forward, once hospitals organise image reading, reporting and cross‑institutional collaboration around this platform, the software may become central to clinical operations.
This pathway has delivered tangible commercial outcomes. In fiscal2026, Pro Medicus recorded approximately AUD261.7million in revenue and AUD127.5million in operating cash inflow, with disclosed minimum contracted revenue of around AUD1.3billion over the next five years. The market assigns it a valuation well above its current‑period revenue. Long‑term contracts, workflow‑embedded products and proven cash‑generating capacity underpin this valuation.
Melbourne demonstrates the first key insight: in high-frequency, trust-reliant healthcare settings, the daily access portal used by clinicians is itself a valuable asset. Yet smooth image visualisation only addresses operational needs. To interpret what imaging findings mean for an individual patient, clinicians also require the patient’s diagnostic, examination and treatment context.
Tempus’ Data in Chicago: Connecting Single Tests to a Patient’s Lifelong Care Journey
Tempus offers a second benchmark. Its core revenue comes from molecular diagnostics – understanding disease via genetic and molecular information – linking test results, clinical records, data applications and pharmaceutical R&D needs. Its implications for the industry are that value can extend beyond diagnosis when textual data connects a patient’s long‑term disease trajectory, treatment decisions and real‑world outcomes.
In Q2 2026, Tempus generated USD382.5million in revenue, including USD289.3million from diagnostics and USD93.2million from data and applications. Tempus also disclosed delivery of the first version of its oncology foundational model to AstraZeneca. Tempus’s valuation reflects data generated by its diagnostic business, capabilities for deploying data to clinicians and pharmaceutical companies, and its sustained revenue growth.
Diagens Tech’s Foundational Model in Hangzhou: Turning Clinicians from Users into Creators
The promising future of medical AI emerges from a recent in-depth conversation between Dr. Ning Song, Founder and Chairman of the Board of Diagens Tech and Kevin Kelly, founding editor-in-chief of WIRED, the renowned global technology publication. Dr. Song drew an analogy between print shops and office productivity software. Document creation once required outsourced typesetting and printing, a slow and costly process. After document-editing tools became accessible on personal computers alongside printers, the volume of documents people could create exploded. Yet global healthcare still operates differently: vendors select disease areas, build AI tools independently and hand finished products to hospitals. Spending RMB 300 million over three years to build one disease-specific imaging AI tool remains common. Dr. Song envisioned a comparable shift: radiologists can leverage state-of-the-art imaging foundational models to train custom diagnostic AI rapidly, accurately and affordably for their own specialties. He described this future medtech ecosystem as “a flourishing bloom of custom clinical AI solutions”. While endorsing this innovative framework, Kevin Kelly noted that for a world-leading medical AI company, the journey to realise this vision forms its strongest competitive moat.
Medical imaging foundational models operate on a simple principle: learn general anatomy and lesion characteristics from vast libraries of CT, ultrasound and pathology scans, then adapt to targeted clinical assessments using specialty-specific patient cases. The shared foundation absorbs the heavy upfront investment, allowing new use cases to reuse existing capabilities. Diagens Tech discloses its iMedImage® foundational medical imaging model spans 19 imaging modalities and 26 clinical specialties. For certain low-data scenarios, a specialty model can be trained within two to three months using roughly 200 imaging samples. Its iMedMaaS® model training and deployment platform enables clinicians to run no-code training and on-premise deployment using in-hospital datasets. Hospitals thereby shift from purchasing off-the-shelf algorithms to defining clinical challenges and co-building tools.
For this production system to work, image volume is only the starting point; clinical expertise must be integrated into datasets. Diagens Tech reports that iMedImage® was trained on more than 80 million medical images. Its iMedLoop™ medical imaging AI R&D and production acceleration platform hosts 28.95 million annotated records, built with contributions from over 3,000 specialised professionals. A “record” refers to one sample entry, which may contain multiple images – much as a single CT scan generates sequential slices. Raw scans expose the model to diverse anatomical variations, while expert annotations identify lesion locations and classification rules. Annotation, quality control and evaluation transform raw image data into usable clinical knowledge for model training. Hospitals can also train models locally using their own data. As of mid-2026, Diagens Tech had partnered with 99 hospitals to deliver 158 model projects. New clinical challenges defined by hospitals, paired with annotation and evaluation frameworks co-developed by specialists, form building blocks for future applications. Notably, the number of models Diagens Tech trains within a single year exceeds the total volume of medical imaging AI models developed across China in the past decade.
Translating this production capacity into clinical adoption requires market-ready products and paying customers. AI AutoVision®, built on iMedImage®, is software designed to help clinicians analyse karyograms and detect abnormalities. It obtained Class III medical device registration from China’s NMPA in May 2026. In its clinical trial, the “clinician+AI” workflow delivered readings in an average of 11.3 minutes, compared with 34.1 minutes for conventional workflows. In the first half of that year, Diagens Tech generated RMB 94.541 million in model service revenue via model and technology licensing, cloud services and on-premise deployment, marking year-on-year growth of 101.1%. The registered product demonstrates how foundational models translate into clinical use, while model service revenue confirms customers are paying for model generation and delivery capacity. The vision of “flourishing bloom” therefore gains commercial substance: clinicians define new clinical challenges, and the platform continuously delivers specialty capabilities using a unified toolset.
The Trillion-Dollar Question: From an Imaging-AI Factory to Disease-Interpretation Infrastructure
What challenges does this “factory” need to tackle? Medical imaging diagnostics cover 14,000 indication-specific tasks. This count refers neither to hardware nor individual images, but to distinct examinations designed for separate clinical objectives. Even for chest CT, lung cancer screening, pulmonary embolism detection and pneumonia assessment represent separate indications, requiring different imaging protocols and clinician interpretation workflows. Each task demands specialised capabilities; medical imaging forms a vast, fragmented landscape of clinical requirements.
Scale brings its own constraint: many clinical conditions lack sufficient patient samples. A study published in Nature Communications identified 5,568 imaging-linked diseases and anomalies across more than 40,000 patient cases, many represented by fewer than 30 samples. The traditional “one model for one disease” paradigm requires fresh case collection, expert annotation, training and validation for every new task. Costs and timelines compound when addressing tens of thousands of niche use cases. Cross-task reusable foundational models embed shared imaging comprehension within a single foundation, allowing clinicians to adapt models using specialty-specific samples. This makes it feasible to cover these long-tail clinical needs at a manageable cost. The true expansion enabled by foundational models lies in the total volume of clinical capabilities that can be built.
This factory unlocks another capability: supporting researchers to frame new questions about disease progression. Imaging captures physical changes within the body. When linked to pathology, genetics, treatment records and patient outcomes, models can probe: which early markers signal recurrence? Which patients will respond best to a given therapy? This is AI for Science in medical imaging: identifying correlations from large-scale clinical observations and generating testable medical hypotheses. Research published in Nature Cancer has sought to infer tumour gene expression from pathology slides to predict therapeutic response.
Diagens Tech’s research advances along this same path. Its iMedImage® technical report reveals the team combined contrast-enhanced CT, clinical and pathological data to predict one-year post-surgical recurrence risk for pancreatic cancer, achieving an AUC of 0.78 across 89 external test samples. An AUC value closer to 1 denotes stronger ability to stratify patients by risk. A separate study combines cervical ultrasound and maternal clinical histories to forecast preterm birth risk. Model research is shifting its focus from “what can be observed today” to “how disease may evolve over time”. In September 2026, Diagens Tech established a joint laboratory with the Hong Kong Polytechnic University to investigate medical imaging analysis, healthcare foundational models and research workflow automation.
The logical thread spanning these three cities now forms a complete loop. Pro Medicus owns the daily software access portal for clinicians; Tempus integrates textual findings into AI-powered long-term patient journey databases for diagnostics and pharmaceutical R&D; Diagens Tech builds a production platform empowering clinicians worldwide to build new models. Its growth follows two clear vectors: expanding into more hospitals, and addressing numerous specialty AI diagnostic needs within each site. Shared foundational models cut development costs for new tasks. Extending beyond diagnosis into prognosis, patient stratification and treatment response research opens opportunities to serve academic institutions and life science companies. This reveals a new dawn for human health management, as the field shifts from Cure Medicine toward Care Medicine.
Amid the promising outlook for human well-being, the three cities, three pathways and three sets of valuations remain underappreciated. Diagens Tech’s growth journey represents humanity’s quest to manage health more effectively and precisely.
Melbourne shapes clinicians’ daily workflows, Chicago maps patients’ lifelong care journeys, and Hangzhou establishes a new benchmark for human health. At the frontiers of AI, we welcome a new era of human well‑being.
