Fig. 2

Constructed image classifier. A Image classifier using the sliding window method. Categorical cross-entropy was used as the loss function, and ADAM software was used to update the parameters. The input layer was set to 96 × 96 with three channels (red, green, and blue), while the output layer was set to two, one containing the cell nuclei and the other without the cell nuclei, respectively. B Image classifier with mask region-based convolutional neural network method (Mask-RCNN). Mask-RCNN consists of three regions: backbone, region proposal network (RPN), and head. The backbone extracts features of the input image. The RPN determines whether each fixed region and the overlap of the regions are correct. The head layer pools the candidate RPN regions to the same image size and then calculates the probability for each class of cytology. CNN, convolutional neural network; ROI, region of interest; Conv, convolutional layer