South Korea Approves Generative Chest X-Ray Report Tool, Bringing Medical AI Into the Text Layer of Interpretation

Chest X-rays are one of the most common imaging examinations in hospitals, and also one of the easiest for backlogs to accumulate. As AI moves from marking suspicious shadows toward writing preliminary interpretation text, what changes is not only a software function, but the daily rhythm in which radiologists work across images, medical records, and reports.

South Korean medical AI company Deepnoid said its generative AI chest X-ray preliminary report tool, M4CXR, has received Class III medical device item approval from South Korea’s Ministry of Food and Drug Safety, with approval number D 제허 26-18호. South Korean media outlet Seoul Economic Daily also reported the approval, meaning the tool has crossed a key regulatory threshold for commercialization in South Korea.

According to Deepnoid, M4CXR can automatically generate preliminary reports for chest X-rays, covering normal images and 41 abnormal findings related to thoracic diseases. The scope listed on the company’s product page includes common clinical clues such as pneumonia, tuberculosis, pulmonary edema, pneumothorax, pleural effusion, cardiomegaly, and fractures; its positioning is not to diagnose patients on their own behalf, but to help physicians form reviewable report drafts more quickly amid high-volume imaging work.

This type of tool differs from earlier medical AI systems that “detect a single lesion.” Deepnoid’s technical materials describe M4CXR as a multimodal medical imaging foundation model for chest X-rays, learning jointly from images and radiology reports, with the goal of pushing image interpretation closer to the level of clinical narrative. The company says the model was trained on more than 10 million pieces of chest X-ray-related clinical data, can process more than 41 findings in a single inference, and generates preliminary results in about 2.3 seconds on average.

Public academic materials also provide some technical background. A 2024 M4CXR paper posted on arXiv positions it as a multimodal large language model for chest X-ray interpretation, with tasks including medical report generation, image grounding, and visual question answering; the paper also says the model can use a prompt approach similar to step-by-step reasoning to first identify imaging findings and then generate a report, and that it can adapt to report generation scenarios involving a single image, multiple images, and multiple examinations. However, the differences between the preprint and the productized version still need to be further clarified through regulatory documents and clinical-use data.

In terms of clinical validation, Deepnoid said M4CXR underwent a multicenter, retrospective, confirmatory clinical trial, with results showing that its report performance achieved non-inferiority compared with radiologists’ reports, and that it maintained consistency across institutions, age groups, and settings including outpatient care, emergency care, health checkups, and hospitalization. This information shows that regulatory review did not look only at the model’s demonstration capabilities, but the publicly available summary still does not fully present the trial scale, endpoint definitions, error types, or performance gaps across different disease subgroups.

The real clinical questions will emerge after deployment. If a generative system can reduce repetitive writing, it may indeed ease pressure on radiology reporting; but it may also produce descriptions that are fluent in tone yet imprecise, or amplify physicians’ review burden in rare, complex, or poor-quality imaging cases. Approval itself therefore does not equal automatic trust, and hospitals still need to design clear human-machine divisions of labor, quality monitoring, and accountability processes.

In November 2025, M4CXR had already been described by Deepnoid as South Korea’s first generative AI-based innovative medical device. Now that it has received item approval, generative medical AI has moved from demonstrations and papers closer to everyday radiology. The key in the next stage is not whether it can write a decent report, but whether, when reports enter real medical record systems and face different hospitals and different patients, it can steadily help physicians complete interpretations faster and more accurately, rather than merely repackaging new review work as efficiency.