Raise a glass to oncology researchers - they just pulled off something remarkable: they took the fast-growing pile of FDA-authorized artificial intelligence and machine learning devices and asked the question every medic, nurse, doctor, patient, and mildly suspicious spouse should ask: “Okay, but what evidence do we actually have?” That may not sound flashy, but neither does checking the brakes before driving downhill. Still a solid move.
AI in cancer care has been arriving fast. Some tools help read scans. Some assist radiation therapy planning. Some flag findings that might otherwise hide in the visual weeds. In theory, this is great. Cancer care is complicated, imaging workloads are heavy, and clinicians are human beings who occasionally need lunch, sleep, or at least coffee that has not been reheated four times.
But the new study, titled FDA-authorized oncology artificial intelligence and machine learning devices and their clinical evidence: A cross-sectional analysis, adds a needed reality check. The authors looked at FDA-authorized AI/ML-enabled medical devices with oncology-specific indications from January 12, 2021, through September 12, 2025. Out of 1,008 FDA-authorized AI/ML devices, 149 had oncology-specific indications. That is about 15%.
So yes, cancer AI is not some distant sci-fi concept anymore. It is already in the building. The more interesting question is whether it has brought enough paperwork, testing, and grown-up accountability with it.
Where Cancer AI Is Showing Up
The study found that oncology AI devices were heavily concentrated in two areas: radiology and radiation oncology.
Radiology accounted for 69 of the 149 devices, or 46%. Radiation oncology accounted for 57 devices, or 38%. That makes sense. AI loves images the way paramedics love a clear house number at 3 a.m. Imaging produces huge amounts of structured visual data, and cancer care leans heavily on scans for detection, staging, monitoring, and treatment planning.
Radiation oncology is another natural fit. Planning radiation therapy involves complex calculations, anatomy mapping, tumor targeting, and avoiding healthy tissue as much as possible. If you have ever watched a radiation plan come together, it feels less like “point beam at bad thing” and more like three-dimensional chess played inside someone’s torso.
The tools in this review were not all doing the same job. Some were related to screening. Some supported diagnosis. Others played a role in treatment. The researchers also examined whether devices fit into the FDA’s computer-aided detection, diagnosis, and triage taxonomy, often shortened to CAD.
That distinction matters because not all AI tools carry the same clinical weight. A device that quietly improves workflow is different from one that directly influences whether a clinician spots a possible tumor or changes a treatment decision. One is like a helpful clipboard. The other is closer to an assistant coach yelling, “Throw the challenge flag.”
The Evidence Question
Here is the number that should make everyone sit up a little straighter: among the 149 oncology-specific AI/ML devices, 113, or 76%, reported clinical testing using patient-derived data.
That is encouraging. Patient-derived data means the device was tested using data from real patients, not just simulated examples or tidy lab exercises. Real-world medical data is messy. Images vary. Patients move. Tumors do not read textbooks. Anyone who has worked in healthcare knows the human body treats “standard presentation” as a polite suggestion.
But the study found much less public evidence for two stronger forms of evaluation.
Only 31 devices, or 21%, reported clinician-in-the-loop testing. That means studies assessed how clinicians performed with the device output. This is a big deal because AI does not practice medicine in a vacuum. A tool may look accurate on paper, but the real test is how it changes human decision-making.
Does it help clinicians find cancer faster? Does it reduce missed findings? Does it create false confidence? Does it add noise? Does it make a radiologist’s day smoother, or does it become one more dashboard blinking for attention like a microwave that wants emotional support?
Then there is prospective testing. Only 7 devices, or 5%, reported prospective testing.
Prospective studies look forward, following performance as data comes in, rather than only looking back at existing datasets. Retrospective evidence can be useful, but prospective testing often gets closer to how tools behave in actual clinical flow. In plain English: it is one thing to review game tape. It is another to play under the lights while the clock is running.
CAD Devices Had Stronger Evidence
One of the study’s most interesting findings was that higher-tier evidence was more common among CAD devices.
The researchers defined higher-tier evidence as clinician-in-the-loop testing and/or prospective testing. Among CAD devices, 20 of 43, or 47%, had this level of evidence. Among non-CAD devices, only 11 of 106, or 10%, did. The difference was statistically significant, with p < 0.001.
That is a pretty loud signal.
It suggests that devices designed to directly aid clinician interpretation may be getting more robust evaluation, at least in publicly available FDA decision documentation. That makes intuitive sense. If a tool is helping a clinician detect, diagnose, or triage cancer findings, the stakes are higher. You want to know how it performs when placed in the clinical decision chain, not just how shiny its algorithm looks in isolation.
This is where risk-based thinking matters. A back-office automation tool and a diagnostic support tool should not necessarily face identical evidence expectations. But when AI starts nudging decisions that affect screening, diagnosis, or treatment, the bar should rise accordingly.
That is not being anti-innovation. That is wearing a seatbelt.
Why This Matters for Patients
For patients, FDA authorization can sound like a finish line. The device has been reviewed. It is allowed on the market. The grown-ups have looked at it.
And yes, FDA authorization matters. But this study reminds us that authorization does not mean every device has been tested in the same way, with the same type of clinical evidence, or in the same real-world conditions.
If AI is helping with cancer care, patients deserve confidence that it has been studied in ways that match its role. A tool that supports treatment planning should be evaluated differently than a tool that organizes images. A device that directly affects a clinician’s interpretation should ideally be tested with clinicians actually using it.
From my old paramedic brain, this lands in a familiar place: equipment is only as good as how it performs in the hands of real people under real conditions. A monitor can have beautiful specs, but if the leads fall off every time the ambulance hits a pothole, congratulations, you now own an expensive rectangle.
Healthcare AI has its own version of that problem. An algorithm can perform well in a dataset and still stumble when faced with different hospitals, scanners, workflows, patient populations, or clinician behaviors.
The Bigger Picture
This study does not say oncology AI is bad. It also does not say these devices are unhelpful. What it says is more specific and more useful: FDA-authorized oncology AI/ML devices are increasingly present, especially in imaging and radiation oncology, but publicly described clinician-in-the-loop and prospective evaluations remain uncommon overall.
That is the sort of finding that should shape policy, research, and hospital purchasing decisions.
Hospitals considering AI tools should ask what kind of evidence supports the device. Was it tested on patient-derived data? Were clinicians studied while using it? Was the evaluation prospective? Does the evidence match the device’s intended role and risk?
Developers should expect more questions, especially for tools that directly shape decisions. Regulators may also continue refining expectations so that evidence requirements fit clinical function, rather than treating all AI tools as if they live in the same bucket.
And clinicians should stay involved. AI in oncology should not be a black box dropped into the workflow like a mystery casserole at a potluck. The people using these tools need to understand what they do, what they do not do, and where the evidence is strong or thin.
A Useful Checkpoint, Not the Final Word
The best part of this study is that it gives the conversation some numbers. Not vibes. Not hype. Numbers.
149 oncology-specific FDA-authorized AI/ML devices. Most clustered in radiology and radiation oncology. Three-quarters with clinical testing using patient-derived data. About one in five with clinician-in-the-loop testing. About one in twenty with prospective testing. Stronger evidence more common among CAD devices.
That snapshot helps separate the useful excitement from the confetti cannon.
Cancer AI may become a major force in earlier detection, better treatment planning, and more efficient care. But if we want it to earn trust, the evidence has to keep pace with the marketing. Especially in oncology, where decisions can carry enormous emotional and clinical weight, “the algorithm said so” is not enough.
The future may involve AI helping clinicians spot subtle findings, plan radiation more precisely, and manage crushing imaging workloads. That future is worth building. But it needs strong evaluation, clear documentation, and a healthy refusal to be dazzled by buzzwords wearing a lab coat.
This blog post discusses research findings and should not be taken as medical advice. If you have concerns about cancer screening, diagnosis, or treatment, please consult a healthcare provider. Research discussed here represents ongoing scientific investigation and clinical validation is still in progress.
All images used in this post are decorative illustrations only and do not represent or reflect the accuracy, reality, or correctness of the referenced research.
Primary Source: FDA-authorized oncology artificial intelligence and machine learning devices and their clinical evidence: A cross-sectional analysis. PubMed Record ID 42025919. https://pubmed.ncbi.nlm.nih.gov/42025919/