Skip to main content
. 2021 Jun 10;12:3541. doi: 10.1038/s41467-021-23913-3

Fig. 2. Obtaining large-scale annotation dataset for deep learning.

Fig. 2

a Representative images of 24 new classifications according to the TBS standard (C1–C24 are the classification indexes during the learning). b Manually annotated 2000 images for each classification. c Training single-class YOLOv3 detection models for 24 classifications. d 81727 smears were detected by 24 single-class YOLOv3 detection models, and the detected targets were manually confirmed. e Cytologists calibrated the ROI of right targets using annotating software, and then the annotated dataset with the corrected ROIs were obtained. (Representative images from calibration process. Scale Bars: 50 μm). f Three cytopathologists confirmed the annotation data again. (Representative images from confirmation process. Scale Bars: 50 μm). g The annotation number of new 24 classifications (totally ~1.7 million annotated targets).