Development of data labeling techniques for terahertz image-based AI cancer diagnosis

Abstract

To develop an AI-based cancer diagnosis technology using terahertz medical imaging, it is imperative to obtain accurately labeled pathological tissue-stained images that serve as the ground truth. However, acquiring these reference images presents challenges such as the cooperation of pathology specialists, high-capacity Whole Slide Images (WSI), and accuracy issues. This study addresses these challenges by creating a brain tumor animal model to obtain tissue-stained images of brain tumor samples, and developing an algorithm using a U-Net neural network for automatic cancerous area segmentation. Through data preprocessing and AI training, high-accuracy labeled data can be provided for terahertz-medical imaging-based AI cancer diagnosis.