陰影画像からのミトコンドリア抽出における教師画像が与える影響の考察

抄録

Organelles play essential roles in cellular function, but conventional fluorescent labeling limits observation to one organelle at a time. Super-resolution shadow imaging provides marker-free visualization of multiple organelles simultaneously. This study focuses on extracting mitochondria from shadow images using deep learning–based segmentation. To evaluate how training data influence performance, we prepared different types of ground-truth images and compared their impact on extraction accuracy. Results show that annotation differences substantially affect segmentation quality, particularly at ambiguous boundaries. These findings highlight the importance of appropriate ground-truth design for reliable organelle extraction from shadowbased imaging.