Lunit Expands Medical AI Across Hospital Operations and Cancer Biomarkers
TECHWORLD ·
✦ AI Summary
Lunit completed the Ministry of Science and ICT's 'AI-specialized foundation model' project and unveiled 'LearningUnit 2.0,' a medical foundation model.
LearningUnit 2.0 was designed at a scale of 30 billion parameters, with around 5 billion actually activated parameters, and was made lightweight enough to run inside medical institutions.
Lunit released it as open source on Hugging Face and also presented the results of a joint AI biomarker study at SITC 2026.
BizCrushWant to follow Korean talks and live streams in real time?Try BizCrush Live
Lunit is expanding the scope of its medical AI applications. The company completed the Ministry of Science and ICT's 'AI-specialized foundation model' project and unveiled 'LearningUnit 2.0,' a medical foundation model developed for healthcare. Lunit announced the development of LearningUnit 2.0 on the 6th.
LearningUnit 2.0 was designed at a scale of 30 billion parameters. However, the number of actually activated parameters is around 5 billion. As a result, LearningUnit 2.0 was made lightweight enough to run inside medical institutions.
A total of 23 industry-academia-research-hospital participating organizations took part in the project. Lunit led the consortium for about 10 months. The company carried out model development and also conducted field validation in clinical settings.
Lunit released the development results as open source on Hugging Face. It also presented the results of a joint AI biomarker study at SITC 2026, the annual meeting of the Society for Immunotherapy of Cancer.
LearningUnit 2.0 was designed with the goal of supporting actual hospital workflows beyond medical question answering. The system is structured to support clinical decision-making, emergency room triage, multidisciplinary care, adverse drug reaction evaluation, and insurance review and claims processing, while linking follow-up tasks based on data analysis and reasoning support.
To verify this setup, field validation was conducted for about 6 months at 9 medical institutions nationwide, including National Health Insurance Service Ilsan Hospital, Kyungwoon University Hospital, and Korea University Medical Center. In the multidisciplinary care evaluation, 86.2% of the treatment recommendations suggested by AI were assessed as clinically valid. In emergency room severity triage, the rate of agreement within one level of specialist judgment was 98%, and in adverse drug reaction causality assessment, the rate of agreement within one level of specialist judgment was also 98%. In insurance review and claims evaluation, 85.6% of the model's answers were rated as usable in practice.
Lunit said that, along with these individual field validation results, it also achieved target performance on major benchmarks, including medical knowledge reasoning, citation factuality, and stability in medical conversations.
Lunit is expanding AI-based cancer biomarker research. In parallel with model development, the company is broadening the scope of AI biomarker research, and it will present the results as two global joint studies using Lunit Scope at SITC 2026, which is being held from the 4th to the 8th in Phoenix, Arizona.
In one of these, a joint study with Labcorp analyzed tumor tissue from 339 ovarian cancer patients to examine differences by whether they had homologous recombination deficiency (HRD). The analysis found that the density of dividing cells within tumors in the HRD patient group was about 1.5 times higher than in the HRP patient group. In addition, the tumor mutational burden (TMB) in the HRD patient group was 6.2 per 1 Mb, while the tumor mutational burden (TMB) in the HRP patient group was 3.1 per 1 Mb, confirming that the TMB in the HRD group was about twice that of the HRP group.
The researchers then applied an AI model that predicts HRD status using only H&E images. As a result, the HRD score produced by the AI showed a statistically significant correlation with the genomic instability score (GIS) measured by genetic testing.
Lunit conducted a joint study with CellCarta and analyzed tissue before and after immunotherapy in patients with muscle-invasive bladder cancer. The study compared differences in immune cell distribution and the distance between immune cells and tumor cells in responders and nonresponders before and after immunotherapy. The analysis found that in the responder group, T cells tended to be adjacent to tumor cells, while in the nonresponder group, immune cell infiltration was lower. In addition, the distance between T cells and tumor cells was relatively greater in nonresponders.
Lunit said these findings confirmed the potential link between immunotherapy response and not only the number of immune cells but also their location and spatial structure. Based on this, the company is planning ways to use AI-based tissue analysis to quantify the distribution of cancer cells and immune cells, as well as the spatial relationship between them. It is also pursuing ways to use these quantification results to predict treatment response.
At the same time, Lunit CEO Seo Beom-seok said that LearningUnit 2.0 is a model for supporting hospital operations overall, rather than being limited to a specific disease or a single function. He also said the company is pursuing the advancement of medical AI technologies that can be used in clinical care and research, based on field validation and global joint studies.
Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=409369
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.
Keep reading
BizCrush