Development of a Surgery-Specialized Generative AI That Brings a "Brain" to the Operating Room—Successful Demonstration in Surgical Settings, Full-Scale Launch Toward Commercialization—

Direava Co., Ltd., as part of the "GENIAC (Generative AI Accelerator Challenge)" project implemented by NEDO and the Ministry of Economy, Trade and Industry (METI) to strengthen generative AI development capabilities, has worked on the "Development of a Vision-Language Integrated AI Foundation Model for Surgical Support" (hereinafter referred to as this project). The company has now developed Japan's first surgery-specialized generative AI (Surgical VLM) capable of understanding surgical situations in real-time and engaging in dialogue.

Conventional generative AI used in the surgical field has been limited to assisting with image recognition (the "eyes") for identifying patient organs and affected areas. The newly developed surgery-specialized generative AI is a groundbreaking system that functions like a "surgeon's brain," understanding "what to do next" based on the surgical situation and engaging in high-level dialogue with the surgeon.

On February 20, 2026, a demonstration test was conducted during an actual gastric cancer surgery at Keio University Hospital. The system successfully confirmed its utility by meeting the required standards for medical education (anatomical accuracy, clinical utility, and textual fluency).

In recent years, the medical field has faced the challenge of training young personnel to inherit advanced medical techniques. Moving forward, Direava will begin full-scale preparations for commercialization (service launch) within 2026 and will proceed with demonstrations for procedures other than gastric cancer surgery. Through this, the company will strongly support the training of surgeons and advanced surgical education, contributing to the future improvement of medical quality and the assurance of safety.

1. Background Since fiscal year 2024, NEDO, in collaboration with METI, has been promoting the GENIAC project to accelerate research, development, and support for generative AI. GENIAC provides support for the development of foundation models ranging from general-purpose models to various domain-specific models. As of the end of March 2026, a total of 53 companies have been developing generative AI foundation models.

Direava, which achieved these results, is one of these companies and is a medical startup originating from the Keio University School of Medicine. The company has been developing medical AI and systems using it, such as the AI-equipped program medical device "Surgical Video Recognition Program Kinosura"*1, which received manufacturing and marketing approval from the Minister of Health, Labour and Welfare on December 9, 2025.

Selected for GENIAC in July 2025, the company developed the "Surgical VLM," Japan's first*2 surgery-specialized generative AI with unprecedented functions, by training its proprietary foundation model on a large dataset of intraoperative images and natural language captions. It then demonstrated these results in an actual surgical setting. In evaluating the developed AI, with the cooperation of Keio University Hospital, the company established criteria considered important not only for accurately grasping surgical situations but also for training surgeons.

[Image: Evaluation items and performance of the developed surgery-specialized generative AI]

2. Results of this Project (1) Development of a Generative AI with a "Brain" that Understands Surgical Context Until now, the use of generative AI in the surgical field has focused on "anatomical recognition (the function of eyes)" to identify patient organs and blood vessels through image recognition. However, "advanced situational understanding (the function of a brain)"—the ability to understand complex surgical procedures and the condition of the affected area, and to reason and converse about "what to do next" or "what to be careful about"—remained unimplemented, posing a major challenge for such AI.

Against this backdrop, Direava worked on developing a surgery-specialized generative AI that meets the standards required by medical education settings. On February 20, 2026, the company operated this AI in an actual surgical environment and completed a demonstration test.

The "Surgical VLM" developed this time is a vision-language integrated AI foundation model that has been trained on a large dataset of high-quality intraoperative image and structural description pairs. It recognizes anatomical structures and surgical situations by taking intraoperative images as input and automatically generates explanatory text in Japanese.

In this project*3, by training the generative AI on vast amounts of data from medical settings, Direava succeeded in building a model capable of grasping the progress of surgery in real-time and conversing with surgeons and medical students. This is a function not found in existing commercialized surgical AI systems, representing the successful development of a world-class, surgery-specialized generative AI.

(2) Demonstration Test at Keio University Hospital To verify the clinical utility and educational effectiveness of the "Surgical VLM," a demonstration test was conducted in an operating room at Keio University Hospital on February 20, 2026. With the cooperation of the university hospital, evaluation items deemed necessary for surgical education were set, and the test verified whether the system could meet a threshold of 80% or higher in anatomical accuracy, clinical utility, and textual fluency. The test results showed 84.7% for anatomical accuracy, 82.9% for clinical utility, and 97.4% for textual fluency, meeting the target values. This confirmed that the developed generative AI accurately recognizes the progress of actual surgeries and demonstrated its practical utility in medical education, such as providing appropriate information required in the operating room and engaging in natural dialogue using professional surgical terminology.

[Image: Demonstration in the operating room]

3. Future Plans Direava will continue to promote research, development, and demonstrations toward the commercialization of the surgery-specialized generative AI developed this time. Aiming for commercialization (start of service provision) within 2026, the company will conduct additional demonstration experiments to expand the range of treatable cases and improve system operability.

Furthermore, the surgery-specialized generative AI developed this time has been trained exclusively on high-quality medical data supervised by Japanese specialists, making it a safe "Japan-originated specialized model" that minimizes hallucinations*4, which are characteristic of generative AI. Through the social implementation of this model, the company will contribute to the inheritance of Japan's world-class "surgical techniques" and the advancement of surgical education while ensuring the security of medical data, ultimately contributing to the improvement of safety across the entire medical field in the future.

[Notes] *1 Program Medical Device "Surgical Video Recognition Program Kinosura" This product is a surgical video recognition program for robot-assisted esophageal malignancy resection, which received manufacturing and marketing approval from the Minister of Health, Labour and Welfare on December 9, 2025 (Approval Number: 30700BZX00326000). The AI assists the physician's perception by detecting and highlighting the left and right recurrent laryngeal nerves in the surgical video in real-time.

*2 Japan's First As of March 2026, according to Direava's research (regarding the development and clinical demonstration of a surgery-specialized generative AI that understands and interacts with surgical situations in real-time in Japan, based on searches of major academic paper databases and press release distribution services).

*3 This Project Project Name: Post-5G Information and Communication System Infrastructure Enhancement R&D Project / Development of Post-5G Information and Communication Systems / Development of a Vision-Language Integrated AI Foundation Model for Surgical Support Project Period: August 4, 2025 – March 31, 2026 Project Overview: Post-5G Information and Communication System Infrastructure Enhancement R&D Project https://www.nedo.go.jp/activities/ZZJP_100172.html#block1

*4 Hallucination A phenomenon where generative AI confidently generates plausible "false" information as if it were fact.

4. Contact Information (Inquiries regarding the content of this news release) Direava Co., Ltd. Representative: Yushi Takeuchi, Representative Director E-mail: info[*]direava.com Website: https://direava.com

NEDO, AI & Robotics Department, Generative AI Team: Endo, Ikeda, Watanabe TEL: 044-520-5241 (General inquiries regarding other NEDO projects) NEDO, Corporate Planning Department, Public Relations Division TEL: 044-520-5151 E-mail: nedo_press[*]ml.nedo.go.jp Please replace [*] with @ when using the email addresses above.

When introducing the organization's name in newspapers, TV, etc., please use "NEDO (New Energy and Industrial Technology Development Organization)" or "NEDO."