Deepnoid's Research Paper Accepted at 'NeurIPS 2025'

Deepnoid, a leading first-generation medical artificial intelligence (AI) company in South Korea, announced on September 23 that its research team's study on chest X-ray image analysis technology, "RadZero," has been accepted at the international AI conference "NeurIPS 2025."

"NeurIPS" is the world's most prestigious academic conference in the field of AI, with approximately 20,000 papers submitted globally each year. Only about 25% of submissions are accepted after a rigorous review process. The selection of Deepnoid's "RadZero" technology demonstrates its international recognition for academic excellence and innovation. The research was conducted solely by Deepnoid's team (Park Jongkwon, Choi Kyoyoon, Yoon Byungmoo, Kim Soobum), highlighting the company's independent technological capabilities and research competitiveness.

The Deepnoid research team will present their findings in San Diego, United States, this December. RadZero is a core technology applied to the multimodal generative AI-based medical device "M4CXR." As a framework equipped with vision-language-based zero-shot multitask capabilities, it breaks down radiology reports into minimal semantic units and precisely matches them with X-ray images.

This enables high-resolution image processing, zero-shot classification of lesions not present in the training data using only text, pixel-level segmentation of lesion areas, and visual grounding that links textual descriptions with image locations.

The most significant innovation of RadZero is its greatly enhanced explainability in AI interpretation. By utilizing its proprietary "Similarity-based Cross-Attention Vision-Language Model (VL-CABS)," it can more clearly explain the rationale behind the connection between text and images. This allows medical professionals to intuitively verify the basis of AI interpretation results, thereby significantly increasing trust in AI technology in clinical settings.

RadZero achieved state-of-the-art (SOTA) performance across classification, grounding, and segmentation tasks when evaluated on eight public benchmark datasets for four types of zero-shot multitask performance.

Without the need for separate labeling work, which typically requires significant time and cost, RadZero can simultaneously perform classification, segmentation, and grounding. It is also capable of handling rare diseases that are difficult to encounter in actual clinical practice, offering a promising supplement to existing medical imaging AI training methods.

Park Jongkwon, head of Deepnoid's AI Advanced Technology Team, stated, "Being accepted at NeurIPS confirms the importance of 'explainability' in medical AI and demonstrates that Deepnoid's vision-language model technology has gained international recognition." He added, "We expect that M4CXR, powered by RadZero technology, will be commercialized in the future and help improve diagnostic efficiency and accuracy in real-world medical settings."

In August, Deepnoid received approval from the Ministry of Food and Drug Safety for the clinical trial plan of M4CXR as a digital medical device. Currently, multicenter, retrospective, and confirmatory clinical trials are underway at Kangbuk Samsung Hospital and Boramae Medical Center to verify the clinical efficacy and safety of M4CXR.