Fig. 1: The diagnostic framework used for classifying SGC and SC.

Training phase: First, the slides from BTH were cutted into tiles (preprocessing module) for training three meaningless tile filter models (meaningless tile filter module) whose weights were then fixed. Then the slides from TMCPLA were also cutted into tiles and all tiles were sifted by the three filter models. Next, CycleGAN were used to produce the tiles with the appearance similar to the tiles from BTH and style similar to the tiles from TMCPLA (style transfer module). Consequently, all tiles passed through the three filter models and the produced tiles from CycleGAN were used to train the models in tile identification module. Finally, the identification results for all tiles were merged to obtain the result for one slide (results merging module). Testing phase: A slide was cutted into tiles and all tiles were sifted through three filter models. All tiles passed through the three filter models were identified by the model in tile identification module. Finally, the identification results for all tiles were merged to obtain the result for the slide.