Soombit AI Trains on 14 Million Medical Images, Pursuing Korea’s First Regulatory Approval for a Generative AI Medical Device
Training vast amounts of medical data for AI models requires both high-level security and cost efficiency. Healthcare AI startup Soombit AI trained on 14 million medical images using Elice Cloud, securing both safety and cost efficiency while pursuing Korea’s first regulatory approval for a generative medical device.Recently, Soombit AI, a healthcare AI startup aiming for Korea’s first regulatory approval of a generative medical device, has been drawing attention. The chest X-ray draft report generation software currently under development by Soombit AI has received approval for its clinical trial plan (IND) from the Ministry of Food and Drug Safety and has entered the clinical stage. The era in which generative AI drafts radiology reports in real clinical settings is fast approaching.Globally, demand for medical imaging exams continues to rise. As screening and early diagnosis become more emphasized, the number of imaging studies such as X-rays, CT scans, and MRIs is rapidly increasing. However, the number of radiologists available to interpret these images is not growing at the same pace. As a result, delays in interpretation, increased workload for medical staff, and delays in care for emergency and critical patients are recurring issues in healthcare settings. Soombit AI is building infrastructure to alleviate these structural bottlenecks with generative AI technology, helping radiologists complete interpretations more quickly and reliably.
Soombit AI’s Vision-Language (VLM)-Based Diagnostic AI ModelSoombit AI is a healthcare AI startup developing a vision-language-based generative AI model that understands both medical images and related clinical information.
Its chest X-ray–focused software, “AIRead-CXR,” automatically generates draft radiology reports based on chest X-ray images. It analyzes a variety of findings that need to be detected in the images and presents personalized draft reports along with information on potential abnormalities, helping radiologists complete interpretations more quickly and consistently.
Radiologists no longer need to write reports from scratch. Instead, they can finalize interpretations by reviewing and editing drafts generated by AIRead-CXR. This approach aims to reduce repetitive documentation tasks and allow physicians to focus more on image analysis and clinical judgment.
Infrastructure Requirements for Safely Training Healthcare DataThe data handled by Soombit AI consists of sensitive healthcare information. Therefore, when selecting a cloud infrastructure, the company’s top priority was not performance or convenience, but data security and regulatory compliance. Because Soombit AI collaborates with domestic medical institutions, it was crucial to confirm whether data centers were physically located in Korea and how storage and computing resources were separated and managed.“As a company dealing with medical data, the data center absolutely had to be located in Korea. At the same time, we had to consider GPU specifications, speed, and pricing. There were not many options that could meet all of these requirements at once.” — Jawook Gu, AI Research Director, Soombit AISolving Security and Cost at the Same Time with Elice CloudSoombit AI chose bare metal servers on Elice Cloud. Training generative models requires large computational resources for each training run, and repeated experimentation is essential to improve model performance. In usage-based pricing structures, costs accumulate every time an experiment is added, which can create psychological and financial pressure when researchers try to expand their experiments. In contrast, bare metal servers provide dedicated resources for a set period, allowing research to proceed by running as many experiments as possible without interruption.“Each training run required substantial computational resources, and we had to run many experiments. In an on-demand environment, adding experiments would continuously increase costs. Thanks to Elice Cloud’s recommendation of bare metal, we were able to expand our experiments with confidence and begin training without hesitation.” — Jawook Gu, AI Research Director, Soombit AIIn addition, bare metal servers allow for completely separated storage structures, which is advantageous from a healthcare data security perspective. With servers located in Korea and storage and computing resources managed separately, Soombit AI was able to build an infrastructure that met regulatory and security requirements more easily.
Training on 14 Million Images and Moving Toward Korea’s First Regulatory ApprovalUsing Elice Cloud, Soombit AI trained a generative medical AI model on 14 million chest X-ray images.
Generative models are highly sensitive to configuration and data composition. To improve performance, numerous combinations must be tested through repeated training and validation. Thanks to the bare metal environment, Soombit AI was able to expand its research more aggressively and steadily improve model performance.“While using Elice Cloud, we were able to run a large number of experiments without burden, and through that process, we improved model performance step by step. As a result, we were able to publish more than four papers at international conferences over the past year.” — Jawook Gu, AI Research Director
▲Soombit AI Published PapersWith sufficient infrastructure and cost efficiency to support extensive experimentation, Soombit AI achieved tangible results not only in academic publications but also in preparing for regulatory approval of a generative medical device. To develop a model suitable for real clinical environments, stability must be verified across diverse patient groups and scenarios. Based on large-scale experiments and training conducted on Elice Cloud, Soombit AI has built the technical foundation to pursue Korea’s first regulatory approval for a generative medical device.
The Next Step Toward Scale-UpSoombit AI’s goal is to build a multimodal foundation model that encompasses not only the chest X-ray generative AI model currently under development but also various imaging modalities such as CT and MRI, along with clinical data.Achieving this requires a cloud infrastructure capable of greater scale and flexibility. As models grow larger and expand into multi-task learning, GPU memory and computational performance become increasingly important, along with storage capacity and speed, and network bandwidth between GPUs and storage. During intensive training phases, high-performance GPU resources must be used heavily, while during idle periods, data archiving capabilities are needed to reduce costs. Stable API communication, network environments, and strong security for service operations are also essential.Elice Cloud ECI is an IaaS platform equipped with these capabilities. It is a container-based cloud infrastructure designed to enable AI companies to carry out the entire process from research to commercial deployment on a single platform. Built on Kubernetes, it allows developers to operate model training and inference services simultaneously in familiar container environments. GPU resources can be dynamically allocated and released as needed, enabling cost-efficient management depending on whether the system is in a training or inference phase.Especially in fields like healthcare AI, where data security and regulatory compliance are critical, being able to leverage container flexibility on top of a stable, Korea-based data center infrastructure certified under CSAP is a major advantage.
An Optimal Cloud Partner for Healthcare AI Companies“We believe it is the most reasonable cloud service operating servers in Korea. Because the data center is domestic, we feel secure about healthcare data protection. With a cost structure suitable for early-stage startups and flexible proposals, we were able to conduct more experiments without burden. That ultimately led to better models and faster results.” — Jawook Gu, AI Research Director, Soombit AIElice Cloud will continue to provide secure, Korea-based infrastructure and efficient AI cloud environments to help healthcare AI companies like Soombit AI develop larger-scale models and services. From early research experiments to commercial service deployment, Elice Cloud aims to build the foundation that allows healthcare AI companies to focus on their technology and products.
1/27/2026
