It is difficult and time-consuming to determine for every pixel whether it is tumor tissue or not.
For decades, the irregular geometry of pleural mesothelioma has quietly undermined medicine's ability to answer the most urgent question in oncology: is the treatment working? A team at the Netherlands Cancer Institute has built an AI model called ARTIMES that measures tumor volume pixel by pixel across CT scans, outperforming experienced radiologists in accuracy and consistency. The achievement, published in Lancet Oncology, does not displace the physician — it sharpens the physician's sight, offering patients earlier answers and sparing them from therapies that have already stopped helping.
- Mesothelioma's sheet-like growth along the lung wall has long made standard tumor measurement criteria nearly meaningless, leaving patients and doctors in a fog of uncertainty about whether treatment is actually working.
- ARTIMES resolves that fog by analyzing every pixel in a CT scan to calculate total tumor volume — a task so granular and time-consuming that no human radiologist can perform it reliably at scale.
- Trained on more than 11,000 scans from 121 hospitals worldwide, the model outperformed experienced radiologists in head-to-head testing, marking the first time AI has demonstrably surpassed physicians at this specific oncological task.
- The clinical payoff is immediate: doctors can now detect treatment failure weeks or months earlier, switch therapies faster, and spare patients from prolonged exposure to toxic and ineffective drugs.
- Access is currently limited to the Netherlands Cancer Institute under EU medical device regulations, but the model is publicly available for research and approval is being sought internationally — with similar AI tools already in development for lung cancer and brain metastases.
Pleural mesothelioma grows not as a discrete mass but as a thin, irregular sheet along the lung wall — a geometry that has long defeated the standard RECIST criteria doctors use to judge whether cancer treatment is working. Built around simple diameter measurements, RECIST was never designed for a tumor that spreads like a film. The uncertainty left patients and physicians unable to say with confidence whether a therapy was helping or failing.
At the Netherlands Cancer Institute in Amsterdam, a team of radiologists, pulmonologists, and AI specialists chose to abandon the diameter approach entirely. Their algorithm, ARTIMES, examines every pixel in a CT scan, distinguishes tumor tissue from healthy lung, and calculates total disease volume — then compares that volume across scans to track whether the cancer is shrinking, stable, or growing. Training the model required more than 11,000 CT images from over 2,000 patients across 121 hospitals worldwide.
When tested against experienced radiologists, ARTIMES won on accuracy, consistency, and reliability. Pulmonologist Sjaak Burgers noted that pixel-by-pixel tumor assessment is simply unfeasible for humans — too difficult, too slow. The findings, published in Lancet Oncology, mark the first demonstrated case of AI outperforming physicians at this task in a way that supports real clinical decisions. Crucially, a physician reviews every measurement before any treatment choice is made; the algorithm amplifies human judgment rather than replacing it.
The practical stakes are high. Earlier detection of treatment failure means stopping an ineffective drug sooner, switching therapies faster, and reducing a patient's exposure to unnecessary toxic side effects. For now, ARTIMES is available only at the Netherlands Cancer Institute under EU medical device regulations, but the team is pursuing broader approval and has released the model publicly for research use. Similar AI tools are already in development for lung cancer and brain metastases — a sign that what began as a solution to one tumor's strange geometry may soon reshape how medicine measures cancer itself.
Pleural mesothelioma grows as a thin, irregular sheet along the lung wall—a geometry that defeats the standard tool doctors have used for decades to measure whether cancer treatment is working. The RECIST criteria, which rely on diameter measurements, were built for tumors that grow as discrete masses. With mesothelioma, the question becomes almost absurd: which diameter do you measure? The uncertainty that follows has frustrated patients and physicians alike, leaving them unable to know with confidence whether a therapy is actually helping.
At the Netherlands Cancer Institute in Amsterdam, a team of radiologists, pulmonologists, and AI specialists decided to solve the problem by abandoning the diameter approach altogether. They built an algorithm called ARTIMES that does something humans find nearly impossible: it examines every pixel in a CT scan, determines whether each one is tumor tissue or healthy lung, and calculates the total volume of disease. Then it compares that volume to previous scans to see whether the cancer is shrinking, stable, or growing. The work required training the model on more than 11,000 CT images drawn from over 2,000 patients across 121 hospitals worldwide.
When the researchers tested ARTIMES against the judgments of experienced radiologists, the algorithm won. It was more accurate, more consistent, and more reliable than human measurement. "This is unfeasible for humans," said pulmonologist Sjaak Burgers. "It is difficult and time-consuming to determine for every pixel whether it is tumor tissue or not." The findings, published in Lancet Oncology, represent the first time researchers have demonstrated that artificial intelligence outperforms physicians at this particular task—and that doctors can confidently base treatment decisions on those measurements.
The practical implications are immediate. When a patient's tumor stops responding to chemotherapy, doctors can now detect that failure weeks or months earlier than the old criteria would allow. That means stopping an ineffective drug sooner, switching to a different treatment faster, and sparing the patient from prolonged exposure to toxic side effects. It also means clearer answers to the question that haunts every cancer patient: Is this working? Kevin Groot Lipman, the technical physician who led the study, emphasized that the model's measurements are always reviewed by a physician before any clinical decision is made. The algorithm is a tool that amplifies human judgment, not a replacement for it.
Currently, only physicians at the Netherlands Cancer Institute can use ARTIMES, a restriction imposed by European Union regulations that govern medical devices developed in-house. But the researchers are already pushing for broader approval. They have made the model publicly available online so that researchers worldwide can begin testing it on their own mesothelioma cases. The institute is simultaneously developing similar AI approaches for lung cancer and brain metastases, suggesting that the mesothelioma breakthrough is just the beginning. Groot Lipman expects the model to "come as a shock" to physicians and researchers outside the mesothelioma field—a signal that artificial intelligence may be about to transform how doctors measure and respond to cancer across many tumor types.
Citações Notáveis
We are the first in the world to demonstrate that AI outperforms humans in this area, and that physicians can actually base their decisions on it.— Kevin Groot Lipman, technical physician and lead author of the study
We can discover the lack of response much sooner than before. This allows a patient to stop treatment earlier and switch to a different treatment, sparing unnecessary side effects.— Sjaak Burgers, pulmonologist at Netherlands Cancer Institute
A Conversa do Hearth Outra perspectiva sobre a história
Why does the shape of the tumor matter so much? Can't doctors just measure what they can see?
With mesothelioma, there's almost nothing discrete to measure. It's not a ball or a lump. It's a thin coating that spreads along the lung wall like frost on a window. The old criteria were designed for tumors you can point to. They don't work here.
So the AI just looks at every pixel and decides what's tumor and what's not?
Exactly. It's trained on thousands of scans where radiologists have already marked what is and isn't disease. Once it learns the pattern, it can do it faster and more consistently than any human could.
But if a computer is making the decision, why do doctors still need to review it?
Because a measurement alone doesn't tell you what to do. The AI says the tumor shrank by 15 percent. A physician has to interpret that in the context of the patient's overall health, side effects, and what other options exist. The algorithm removes the guesswork from measurement. The judgment still belongs to the doctor.
How much earlier can they catch treatment failure now?
Weeks or months. That's the real gift. Instead of waiting for a diameter-based measurement to show obvious growth, you can see the volume isn't changing. You can stop the drug and try something else before the patient has endured months of unnecessary toxicity.
What happens next? Does this stay locked in Amsterdam?
No. They're working on approval to use it elsewhere. And they've already released it publicly so researchers can test it. They're also building similar models for other cancers. This is the proof of concept. Once you show it works for one tumor type, the question becomes: what else can AI measure better than we can?