□ DGIST (President Lee Kunwoo) Department of Robotics and Mechatronics Engineering Professor Park Sang-hyun and his research team will present three papers at the 27th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI).
□ MICCAI is a world-class conference in the field of medical imaging, and this year's conference will cover cutting-edge image computing and machine learning techniques, computer-assisted interventions, and fundamental and innovative research topics on a variety of clinical problems. Prof. Park's team has been consistently publishing papers at the highly competitive MICCAI conference since 2019 and has already published a number of papers this year, making their research noteworthy in the domestic and international academic community.
□ Prof. Park's team will present three full papers at the conference. The first paper is on a multi-instance learning (MIL) technique for detecting and classifying abnormal lesions using large pathology images and disease labels. The study proposes a technique that efficiently finds significant instances using a visual language foundation model and improves performance stability by adding textual instructions. In addition, an adaptor that performs well with less data is introduced to extract features fit for pathological image analysis. This method performed well on a cancer diagnosis model using two pathological images, and its clinical applicability was demonstrated by performing reliable cancer diagnosis with data collected from various hospitals and visualizing important regions related to text.
□ The second paper proposes a technique that automatically segments cell nuclei in pathology images using only point labels. Existing methods require the manual segmentation of a large number of cells, making data construction challenging, but this study addresses the issue by utilizing the Segment Anything Model and point labels to create temporary training labels. An adaptor was added to the image encoder for fine-tuning to reduce the domain differences between the foundation model and the pathological images. This method successfully performed cell nucleus segmentation while maintaining the generalization performance of the model. It also performed better than existing methods and demonstrated good performance even in few-shot and cross-domain situations.
□ The third paper, developed in collaboration with researchers at the University of Pennsylvania, proposes a method for adapting brainwave classification AI models to new subjects by utilizing resting-state brainwave signals. Existing models often struggle with poor performance when applied to new subjects because of interpersonal variabilities; the team addressed this by generating task signals from resting-state brainwave signals and adapting the models with three different loss functions. This method allows the model to easily adapt to new subjects without the need to collect task signals separately. It achieved an average classification accuracy improvement of 8.53% when validated on three datasets.
□ “We are pleased to present three papers at MICCAI, a leading medical imaging conference in the world,” said Prof. Park. “These studies demonstrate how the recently proposed foundation models can be applied well to various subproblems in the medical field. We will continue to work on developing techniques that can perform accurately on various diseases and in different environments and further improve the generalization performance of AI models.”
- Corresponding Author E-mail Address : [email protected]
