AIDOT Inc., a medical artificial intelligence company, participated in Medical Imaging with Deep Learning 2026 (MIDL 2026), an international medical imaging AI conference held in Taiwan, where it presented two medical AI research papers conducted jointly with researchers from leading university hospitals in Korea.

The presentations focused on enabling on-device AI operation in clinical environments with limited infrastructure and improving the detection of small lesions through an understanding of anatomical knowledge.

The first paper presents research on the on-device optimization of Cerviray AI, AIDOT’s cervical cancer screening system. The study was conducted jointly with the research teams of Professors Jaeyoon Song and Sungmin Kim of Korea University Anam Hospital and Professor Donghoon Seo of Seoul National University Bundang Hospital.

To reduce the dependence of conventional deep learning models on high-performance GPUs, the researchers used knowledge distillation to compress a ViT-Base model into a ViT-Tiny model and applied post-training quantization.

As a result, the optimized model achieved device-independent inference at a speed of 3.4 seconds per image using the CPU of a Samsung Galaxy Tab S7.

The findings demonstrate the potential to use AI for cervical cancer screening in low- and middle-income countries and primary healthcare institutions where network infrastructure and high-cost servers may be limited.

The study is also significant because operating AI directly on devices within hospitals can reduce security concerns associated with transmitting data to external servers.

The second paper covers a multi-stage fine-tuning method developed to improve urinary stone segmentation performance in non-contrast CT images. The research was conducted in collaboration with the teams of Professors Young Eun Yoon and Jaehoon Oh of Hanyang University Hospital and Professor Donggeon Lee of Seoul National University Bundang Hospital.

Urinary stones are generally small and occupy only a sparse number of voxels, which limits detection performance when AI models rely solely on learning visual features.

Using the VISTA3D model, the researchers first trained the AI to recognize organs associated with urinary stones and then transferred this anatomical knowledge to the urinary stone segmentation task.

The study showed that enabling AI to understand the surrounding anatomical context can contribute to improving the detection of small lesions.

AIDOT plans to apply the findings of this research to its urinary stone detection solution, URO dot AI, to further improve its urinary stone detection performance.

Hansol Choi, Chief Technology Officer of AIDOT, said, “This research is the result of close collaboration with leading clinical institutions to address practical challenges in medical settings. We will continue conducting research to optimize AI for mobile clinical devices and improve its anatomical understanding, with the aim of advancing reliable AI-assisted diagnostic solutions.”

Korea Economic TV / 2026-07-09 / Deputy Managing Editor Jae-jun Yang