A radiologist can ask: What is the shape? What do the margins look like?
Every year, lung cancer claims lives that might have been saved by earlier, more certain diagnosis — a burden that falls on radiologists tasked with reading ambiguous shadows in chest scans, often alone with their experience and its limits. A research team at Meijo University in Japan has built an AI system that does not merely render a verdict, but converses: answering a clinician's questions about what it sees in natural language, the way one thoughtful colleague might speak to another. In doing so, they have begun to address one of medicine's quieter crises — not the absence of tools, but the absence of tools that can be understood and therefore trusted.
- Lung cancer's lethality hinges on timing, yet diagnostic inconsistency between radiologists means the same scan can yield different conclusions depending on who reads it.
- Conventional AI diagnostic tools compound the problem by delivering binary verdicts with no explanation, leaving clinicians to either accept or reject a machine's judgment without understanding its reasoning.
- The Meijo University team built a visual question answering system that lets radiologists interrogate CT scans in natural language — asking about nodule shape, margins, and texture — and receive clinically meaningful, explainable responses.
- Tested against a large annotated dataset, the system achieved a CIDEr score of 3.896, demonstrating not just technical accuracy but language that mirrors how physicians actually document findings.
- For patients too medically fragile to undergo biopsy, this kind of reliable, explainable imaging diagnosis is not a convenience — it is the only path to knowing.
Lung cancer is among the deadliest of malignancies, and its prognosis turns sharply on how early it is caught. Radiologists search chest CT scans for subtle nodules — assessing shape, texture, and margins — in work that demands deep experience and carries real variability. Two physicians examining the same image may not reach the same conclusion.
Artificial intelligence has grown capable of detecting these nodules, but most systems offer only a binary output: benign or malignant, with no explanation attached. A radiologist receiving such a verdict must decide whether to trust it without knowing why the machine reached it.
At Meijo University in Japan, graduate student Maiko Nagao and her colleagues built something different. Drawing on the Lung Image Database Consortium — a repository of expert-annotated CT scans — they converted structured morphological descriptions into natural language and trained a vision-language model to answer specific clinical questions about what it observes. A radiologist can now ask the system about a nodule's margins or spiculation and receive a response written the way a physician would write it in a report.
The system performed with notable accuracy, achieving a CIDEr score of 3.896 — a measure indicating that its language was not merely correct but clinically coherent and contextually appropriate. Crucially, it remained consistent across different nodule characteristics, a prerequisite for any tool meant to be relied upon.
Nagao's motivation was grounded in a specific clinical reality: patients whose other medical conditions make biopsy too dangerous, leaving imaging as their only diagnostic option. For them, the quality of that interpretation is not a matter of convenience but of survival.
Published in the International Journal of Computer Assisted Radiology and Surgery in early 2026, the work envisions AI not as a replacement for radiologists but as a transparent collaborator — one that shows its reasoning, supports report writing, trains less experienced clinicians, and helps standardize how findings are described across institutions. The ambition is a medicine where machines and physicians think together, visibly and accountably.
Lung cancer kills more people than almost any other malignancy on earth. The difference between catching it early and catching it late is the difference between years of life and months. Radiologists spend their days staring at chest CT scans, looking for the small white shadows that might be tumors—assessing their shape, their edges, their texture, the way they sit inside the lung. It is meticulous work, and it depends entirely on how much experience a radiologist has accumulated. Two doctors looking at the same scan can reach different conclusions.
Artificial intelligence has gotten better at spotting these nodules. But most AI systems do only one thing: they say yes or no, benign or malignant. They offer no explanation. A radiologist sees a result and has to decide whether to trust it, and that decision is made in the dark.
A team at Meijo University in Japan, led by graduate student Maiko Nagao and professors Atsushi Teramoto, Hiroshi Fujita, and Kazuyoshi Imaizumi, decided to build something different. Instead of a black box that spits out a diagnosis, they created a system that talks back. Using a technique called visual question answering, they trained an AI model to look at a CT scan and answer specific questions about what it sees. A radiologist can ask: What is the shape of this nodule? What do the margins look like? Is there spiculation? And the system responds in natural language, the way a physician would write it in a report.
The researchers built their system using data from the Lung Image Database Consortium, a repository of annotated scans where experts had already documented the morphological features of thousands of nodules—their sphericity, their margins, their texture, whether they showed lobulation or calcification or spiculation. Nagao and her team converted these structured annotations into natural language descriptions, then paired them with clinical questions. They fine-tuned a vision-language model on this dataset, teaching it to connect what it saw in an image with what a physician would say about it.
When they tested the system, it worked. The model generated findings that matched what radiologists had written, with a high degree of accuracy measured by a metric called CIDEr score—they achieved 3.896, indicating that the language was not just technically correct but clinically meaningful and contextually appropriate. The system stayed consistent across different morphological features, which matters: a diagnosis is only useful if you can rely on it.
But the real innovation was the interactivity. A radiologist no longer receives a verdict from a machine. Instead, they can ask questions. They can explore the lesion's characteristics in real time, getting detailed, explainable answers that show their reasoning. This transparency changes everything. It makes the AI a tool that a clinician can understand and therefore trust. It can help write reports faster. It can teach younger radiologists what to look for. It can reduce the variability that comes from one person's experience being different from another's.
Nagao was motivated by something she had seen in practice: patients who could not have biopsies because their other medical conditions made the procedure too risky. For them, getting the diagnosis right from imaging alone was not just convenient—it was necessary. She wanted to build a tool that would help doctors make better decisions without putting patients through additional harm.
The work, published in the International Journal of Computer Assisted Radiology and Surgery in March 2026, points toward a different kind of medical AI. Not a replacement for radiologists, but a collaborator. Not a black box, but a system that shows its work. In the near term, it could standardize how nodules are described and interpreted across hospitals. Over time, it could reshape how doctors and machines work together in the clinic, making diagnosis more transparent, more consistent, and ultimately more reliable for the patients who depend on it.
Notable Quotes
Conventional AI diagnostic support methods lacked explainability because they mainly focused on classifying lesions as benign or malignant. Our goal was to generate findings similar to those written by physicians to improve the usability and acceptance of AI outputs.— Maiko Nagao, graduate student, Meijo University
Some patients cannot undergo invasive tests like biopsies because of other medical conditions. By making AI outputs more understandable and interactive, we hope to improve both clinical decision-making and patient care.— Maiko Nagao
The Hearth Conversation Another angle on the story
Why does it matter that the AI explains itself? Couldn't a radiologist just trust the diagnosis if it's accurate?
Because trust without understanding is fragile. A radiologist needs to know whether the AI is looking at the right features, whether it's reasoning the way a human would. If something seems off, they need to be able to ask why.
So this is really about making the AI legible to doctors.
Exactly. And not just legible—collaborative. A radiologist can now ask targeted questions about what they're seeing, almost like consulting with a colleague who never gets tired.
The system was trained on annotated scans. How much does the quality of that training data matter?
Everything. The annotations describe what experienced radiologists saw in those images. The AI learns to connect visual patterns to the language radiologists use. If the annotations are sloppy, the system will be sloppy.
And the patient benefit—is it just that they avoid a biopsy?
That's part of it. But it's also about consistency. If every radiologist interprets a nodule the same way, fewer patients get called back for unnecessary follow-ups, and fewer cancers slip through because someone missed something.
What happens next? Does this go into hospitals?
That's the question. The research shows it works in the lab. Now it has to prove itself in real clinical workflows, with real radiologists, real time pressure, real stakes.