Better Medical AI Starts With Better Medical Data

What Imaginostics reveals about the missing layer between AI models, biological data and precision medicine

Artificial intelligence in healthcare is often presented as a race for better models: more powerful algorithms, larger datasets, increasingly capable multimodal systems and more accurate predictions. That framing begins one layer too late. In medicine, an AI system can only extract value from what the underlying measurement system makes observable. When the biological signal is incomplete, indirect or difficult to quantify, even the most sophisticated model starts with a structural limitation. Better algorithms can extract more from existing data, but they cannot recover information that was never captured.

Some of the most important advances in medical AI are therefore likely to come not only from better models, but from better ways of measuring the human body and converting those measurements into reliable, patient-specific and clinically relevant data. Imaginostics, a Franco-American startup, offers an instructive example. Instead of asking only how AI can analyze existing medical data more effectively, it raises a more fundamental question: what happens when we improve the information given to AI in the first place?

The blind spot in the medical AI race

Most discussions about medical AI begin after the data already exist. We examine model accuracy, training datasets, bias, explainability, regulatory compliance and human oversight. All of these issues matter, but they assume that the information entering the system already represents the biological reality we want AI to understand.

An AI model does not observe a patient directly. It observes a representation created by another system: an MRI or CT scanner, a pathology slide, a wearable sensor, a laboratory test or genomic sequencing. AI sits downstream of an entire measurement architecture that determines what can be observed, how precisely it can be quantified and what remains invisible. The regulatory record shows where the effort has gone: of the more than 1,500 devices on the FDA's list of AI-enabled medical devices, roughly three-quarters are classified under radiology, most of them applying AI to imaging modalities medicine has used for decades.

The intelligence layer cannot indefinitely compensate for limitations in the measurement layer. AI can identify patterns that humans overlook and extract weak signals from complex datasets, but inference is not measurement. If a biological phenomenon has not been captured with sufficient fidelity or specificity, the model is reasoning from an incomplete representation of reality. The future of medical AI is not an algorithm problem alone. It is also a sensing, measurement and data-generation problem.

From interpreting an image to measuring biology

In April 2026, at eMerge Americas in Miami, I attended a presentation by Imaginostics. On the screen were two representations of the same brain. The first looked like what we instinctively associate with an MRI: anatomy, contours and contrast variations for a radiologist to interpret. The second revealed something else: a network of much finer vessels, paired with quantitative measurements of their density, structure and permeability. The change is more than technical. Imaginostics is trying to move imaging from a logic in which clinicians mainly interpret a picture toward one in which certain characteristics of the smallest blood vessels can be measured directly in an individual patient, within a single MRI examination.

Co-founded by CEO Valerie Gharagouzloo, a lawyer turned healthtech entrepreneur, and CSO Dr. Codi Gharagouzloo, an engineer and MRI physics researcher, the company grew out of research developed in Boston. Now based in Orlando, Florida, it is working to move the technology from research toward clinical use. “We transform the MRI signal from an essentially qualitative signal into actionable quantitative data, from which we can extract measurable vascular biomarkers,” Valerie Gharagouzloo explains. The company relies on two complementary platforms designed to operate with existing MRI scanners.

ImagiView is designed to generate quantitative vascular imaging data using an iron-based contrast agent instead of gadolinium. According to the company, the FDA granted the technology Breakthrough Device designation in October 2024 for patients with chronic kidney disease stages 3 to 5 or end-stage renal disease, for whom gadolinium-based agents in MRI and iodine-based agents in CT are contraindicated. The designation gives access to closer interaction with the agency and prioritized review. It is not a marketing authorization, and Imaginostics describes itself as a pre-FDA company that is not yet providing clinical services.

ImagiSight, the company's AI layer, is being developed to work from ImagiView data and extract quantitative vascular biomarkers such as cerebral blood volume, small-vessel density and blood-brain barrier permeability. The product architecture mirrors the argument of this article: measurement first, intelligence second. Information moves from image to quantitative data, from data to vascular biomarker and from biomarker to clinical information. AI matters because it can help turn complex quantitative information into usable insight, but it is not the starting point. The starting point is the ability to measure something more precisely than before.

The measurement ceiling

There is a useful parallel with physical AI. An autonomous system may have an exceptionally capable model, but if its sensors cannot reliably perceive distance, movement or obstacles, intelligence alone cannot produce consistently good decisions. The quality of perception constrains the quality of reasoning. It is also why, as I argued in The Platform War Behind Robotics, the scarce asset in physical AI is shifting from compute to real-world data.

Healthcare follows the same logic. The human body is the physical system, medical instruments are its sensors, and the data they generate form the AI system's representation of biological reality. This suggests what I call the measurement ceiling:

The potential performance of an AI system is constrained by the quality, relevance and reproducibility of the physical measurements from which its data are derived.

Information theory has a name for the mechanism: the data processing inequality, which states that no downstream processing can increase the information a signal carries about its source. Better models, more compute and larger datasets can extract more from the available signal, but they do not raise that ceiling. Improving the measurement layer can. This also separates the argument from data-centric AI, which improves how existing data are labeled, cleaned and curated. The measurement ceiling is set earlier, by what the instrument captures in the first place.

A better architecture for medical AI

For executives, investors and healthcare organizations, this suggests evaluating medical AI as a chain of five layers rather than as a model. The first layer is measurement: what biological phenomenon can the technology actually capture? The second is quantification: can that phenomenon be converted into reproducible numerical information at patient level? The third is intelligence: what additional understanding does AI create from those data? The fourth is clinical decision: does the information change diagnosis, treatment selection, monitoring, prevention or trial design? The fifth is evidence: can the full chain demonstrate scientific validity, clinical relevance, reproducibility, safety and economic value?

Measurement → Quantification → Intelligence → Decision → Evidence

Most medical AI discussions still concentrate on the third layer, intelligence. Healthcare value depends on the entire chain.

From population averages to individual trajectories

The shift toward quantitative measurement becomes especially important as healthcare moves from population-based medicine toward precision medicine. Modern medicine has generated extraordinary knowledge by studying groups, but the person entering a hospital is not a population. The challenge is to reduce the gap between what medicine understands statistically across cohorts and what it can determine biologically about one individual.

ImagiSight aims to characterize vascular health at patient level, including microvascular density, small-vessel integrity and blood-brain barrier leakage. If such measurements can be reproduced over time, medicine gains another dimension: not only comparing a patient with others, but comparing that patient with themselves. The question evolves from How does this patient compare with a population? toward How is this patient's biology changing over time? That shift from static comparison toward longitudinal understanding is central to precision medicine.

It is particularly relevant in neurodegenerative disease. Research has linked blood-brain barrier breakdown to early cognitive dysfunction, independently of amyloid and tau (Nation et al., Nature Medicine, 2019), and Imaginostics is investigating whether its approach can make such changes measurable in individual patients. The approach has so far been demonstrated in preclinical models, including a 2021 study in rats carrying the human APOE4 gene, the strongest genetic risk factor for late-onset Alzheimer's disease. A 96-subject clinical study in mild cognitive impairment and vascular dementia, supported by the Alzheimer's Drug Discovery Foundation and a National Institute on Aging SBIR grant, is designed to test it in humans.

The claim is not that MRI can already predict Alzheimer's disease decades before symptoms appear. The more precise proposition is that better vascular measurement may help researchers characterize biological changes earlier and more accurately than conventional approaches allow today. That distinction is the evidence layer of the chain at work.

Better data could reshape drug development

The same principle applies to pharmaceutical research. Drug development is fundamentally a measurement problem: researchers need to know whether a therapy is changing the biological process it was designed to affect. In many conditions, particularly slowly progressing diseases, clinically visible outcomes can take years to emerge. Quantitative imaging biomarkers could provide additional tools to stratify patients, characterize disease progression and evaluate treatment response, which is why Imaginostics sees pharmaceutical development as an important application for its vascular biomarkers.

Better medical data creates several downstream sources of value. The same measurement technology could help clinicians understand individual patients while helping pharmaceutical companies evaluate therapeutic effects. The strategic asset is the information architecture around the model.

The competitive moat is moving upstream

This matters because the model layer is becoming increasingly competitive. Foundation models are improving rapidly, open-source capabilities continue to advance and performance gaps can narrow quickly. Healthcare companies need to ask where defensible advantage will come from once sophisticated AI becomes broadly available. Part of the answer will lie in clinical evidence, workflow integration and regulatory positioning. Another important source of differentiation will be access to biological information that competitors cannot easily reproduce.

A company applying a differentiated algorithm to widely available medical data and a company generating a new quantitative biological signal can both create value, but their competitive structures are very different. Algorithms can be replicated, improved or displaced. A validated measurement technology combined with proprietary datasets, longitudinal evidence and embedded clinical workflows is much harder to reproduce. The race is shifting from Who has the best algorithm? toward Who has access to the most meaningful biological information?

Better data does not eliminate the trust problem

Better measurement does not automatically produce trustworthy AI. A high-quality biological signal can still feed an inadequately validated model. A quantitative biomarker can be interpreted incorrectly, performance can vary across populations or clinical environments, and AI-generated recommendations can be overtrusted.

Trust has to be earned at every link of the same five-layer chain. The measurement must be reproducible, the data traceable, the model validated for its intended use and clinical responsibilities explicit. Human oversight must be designed into the workflow rather than added afterward, a principle that applies to clinical AI as much as to AI agents acting on an organization's behalf. The objective should not be autonomous medicine for its own sake, but better evidence enabling better human decisions.

From breakthrough to scale

Scientific innovation is only the beginning. Healthcare has one of the most demanding scale-up pathways of any industry because a breakthrough must survive several transitions: scientific validity must become clinical evidence, evidence must support regulatory processes, the product must integrate with existing workflows, clinicians must trust it and health economics must justify adoption.

Imaginostics is navigating precisely this transition. One advantage is that its platforms are designed to operate with existing MRI scanners rather than requiring hospitals to replace installed imaging infrastructure. Access to that installed base still runs through scanner manufacturers, which is why the Letter of Support issued by Siemens Healthineers in March 2026 matters: it opens a dialogue on technical and commercial pathways for MAGNETOM MRI systems, although the announcement is explicit that any future collaboration remains subject to further discussions and definitive agreements.

Yet compatibility alone does not create scale. A more realistic equation is multiplicative, and if one factor approaches zero, the commercial potential of the entire system collapses:

Science × Evidence × Regulation × Integration × Economics

This is also where ecosystem strategy matters. Imaginostics was founded in the United States and has built links with French and European healthcare and innovation ecosystems, including the Paris-Saclay Cancer Cluster and Future4care's Go-To-Market program, illustrating how medical deeptech companies increasingly need to connect research, clinical partners, regulators and markets well before they reach scale.

Five questions leaders should ask

The Imaginostics case suggests five practical questions for anyone evaluating medical AI. They extend far beyond MRI, applying equally to digital pathology, genomics, wearable sensing, remote monitoring and medical robotics.

What new biological information becomes measurable? If a technology only analyzes existing data differently, its differentiation will depend on model performance.

Is the information quantitative and patient-specific? Repeatable individual measurement creates the possibility of understanding biological trajectories rather than relying only on population correlations.

What does AI contribute beyond the measurement itself? “AI-powered” is not, by itself, a value proposition.

Can the complete chain be demonstrated? Measurement quality, data integrity, model performance, clinical validation and human oversight form one system.

Can the technology scale inside healthcare as it exists today? Infrastructure compatibility, workflow integration, economics and adoption often matter as much as technological superiority.

The next frontier of medical AI begins before AI

The coming years will bring more powerful medical AI models, capable of reasoning across images, clinical records, genomics, pathology and longitudinal patient histories. As that intelligence becomes broadly available, the decisive advantage is likely to move upstream. The winners will be the organizations capable of giving those models access to biological information that medicine could not previously measure with sufficient precision: better sensing, better quantitative biomarkers, better longitudinal data and stronger evidence connecting those signals to meaningful clinical decisions.

The Imaginostics case matters beyond one startup or one imaging technology because it illustrates a broader transition in healthcare innovation: from using AI primarily to interpret the medical world as we currently observe it, toward using technology to make that world observable in new ways. The long-term promise of medical AI goes beyond recognizing disease more accurately once it becomes visible. It is to help medicine understand biological change earlier, follow it more precisely and intervene with greater confidence before deterioration becomes irreversible.

The defining question for the next generation of medical AI is therefore not only How intelligent is the model? It is How well can we measure the patient? Before medicine can become truly predictive, personalized and preventive, it has to make human biology more measurable. That is why better medical AI starts with better medical data.

An earlier French version of this analysis was published in Alliancy on 14 September 2026.