Abstract
Objective : To clarify the magnification of microscopy capturing the image data to design the architecture of deep learning-based artificial intelligence (AI) for gynecological cervical cytology.
Study Design : 187 and 264 slide images were captured using ×20 and ×40 magnification, respectively. From the 451 slide images, a total of 996 cell and/or cell clusters images (536 for ×20 and 460 for ×40 images) were individually annotated according to the Bethesda system, followed by trichotomizing into the negative for intraepithelial lesion or malignancy (NILM), low-risk (atypical squamous cells of undetermined significance [ASC-US] and low-grade squamous intraepithelial lesion [LSIL]), or high-risk (atypical squamous cells cannot exclude HSIL [ASC-H], high-grade squamous intraepithelial lesion [HSIL], and squamous cell carcinoma [SCC]) groups for training data set. The training data set was split into 80% and 20% segments as training set and validation set, respectively. AI was established by deep learning using a faster region-based convolutional neural networks (Faster R-CNN) system, and the trained network was evaluated using precision and recall via 5 independent tests.
Results : Overall, the mean precision and recall using ×40 images were inclined to be higher than ×20 images among the three categories, especially representing statistical significances with precision in high-risk (p<0.05) and recall in NILM groups (p<0.05), respectively.
Conclusion : These results established for us the importance of the necessity for good quality of training cellular images to establish diagnostic AI for gynecological cervical cytology.
