Fig. 6

Overview of the dataset creation process. (a) Flowchart for acquiring the training dataset. After converting the LBC specimens to WSIs, they were categorized into tiles, and images with a cell quantity of 30% or more were filtered and selected. (b) The process for acquiring data with a “normal” label using a similar image search. One image without cell overlap and one with overlap were used, and a similar image search was conducted using a color hash. From the resulting image pool, 500 images were each randomly selected. They were labeled as “normal.” (c) The process for acquiring data with an “abnormal” label. Hand labeling was performed for each tile image. If atypical cells were identifiable in the tile image, the label “abnormal” was assigned. (d) Flowchart for acquiring the test dataset. After converting LBC specimens to WSIs, cases that could not be scanned were excluded, and the remaining specimens were categorized into tiles. Images with a 30% or more cell quantity were filtered and selected.