Can generative AI transform medical imaging and diagnosis?

At least two Korean healthcare companies, including Deepnoid, are reportedly preparing to seek approval for generative AI medical devices. In January, the Ministry of Food and Drug Safety released the "world's first" review guidelines for generative AI medical devices. Attention is now focused on why the MFDS created separate review guidelines specifically for generative AI medical devices.

Deepnoid recently held a press conference with reporters covering the Korean medical device industry to showcase the potential usefulness and features of generative AI medical devices. The company has developed a generative AI medical solution, M4CXR (Multimodal for Chest X-ray), which is currently in clinical trials for marketing approval. If M4CXR is approved before its competitors, it will become the "world's first" generative AI medical solution, according to Deepnoid.

M4CXR provides over 80 findings, including nodules, masses, sclerosis, fibrosis, atelectasis, emphysema, tuberculosis, pleural effusion, pneumothorax, cardiomegaly, enlarged lymph nodes, rib fractures, and relief. While existing AI medical devices only identify issues, M4CXR goes a step further by offering solutions (diagnoses) to those problems.

According to Deepnoid, M4CXR's features include automatic drafting of chest X-ray readings, consistent and reliable interpretations using AI trained on diverse datasets, reduction of repetitive work hours, faster and more accurate diagnoses when needed, and streamlined workflow with AI models trained on vast amounts of chest X-ray data.

Choi Hyun-seok, a radiologist and chief medical officer (CMO) at Deepnoid, stated that M4CXR is "as efficient as having a first-year trainee doctor reading alongside you."

“Radiologists are good at reading without AI, but M4CXR will help non-specialists get closer to radiologists,” Choi said. “This may lead to an increase in the number of radiology specialists.”

Deepnoid stated that it has confirmed the promise of M4CXR through technical and clinical research papers. In a technical study titled “M4CXR: Exploring Multi-task Potentials of Multi-modal Large Language Models for Chest X-ray Interpretation,” M4CXR improved CXR interpretation, demonstrated higher clinical accuracy and multi-tasking capabilities compared to existing models, and validated its performance in medical report generation, visual grounding, and question-and-answer tasks.

In addition, a clinical study titled “Comparative Analysis of M4CXR, an LLM-based Chest X-ray Report Generation Model, and ChatGPT in Radiological Interpretation” evaluated the diagnostic accuracy and clinical utility of M4CXR and ChatGPT, focusing on chest X-ray reading performance and exploring their potential applications in clinical settings. The results of the study showed that M4CXR consistently outperformed ChatGPT in several aspects of diagnostic accuracy. However, further validation of its performance in various pathological conditions and additional research on human-AI interaction are still needed.

“The comparison study with ChatGPT was an unfair game because ChatGPT is not designed for medical use. There are still areas that need to be improved,” said Choi. “But the important thing is that we have confirmed the usefulness of M4CXR. Unlike university hospitals that have fully equipped radiology departments, it is clear that M4CXR helps fill the gap in tertiary and secondary hospitals that lack radiology specialists.”

However, critics point out that there are still some downsides to AI image reading. For example, the reading time may increase, and there are concerns about potential reading errors.

In response, Deepnoid Senior Managing Director Kim Tae-gyu said, “Reading chest X-rays is not easy and requires expertise, and since many images are taken in hospitals, universality must also be considered.” He added, “Deepnoid has a lot of know-how in chest imaging through its DeepChest solution. Based on this, it is clear that generative AI can assist medical staff by drafting readings and supporting them in their reading tasks.”

Clinical trials are underway at "multiple healthcare organizations" to ensure the clinical utility and safety of generative AI solutions like M4CXR, allowing them to be actively utilized in clinical settings, Kim added. The company plans to publish the results of these studies later this year.

"We already have evidence that AI medical solutions can reduce reading time and minimize reading errors," Choi said, adding, "There are always concerns about new technologies. If we don’t challenge the development of new technologies out of fear, there will be no progress in healthcare. Of course, we recognize that safety is important. We will continue to validate its efficacy and safety so that clinicians can trust and utilize it."