Fig 1.
Three types tile images we might overlook.
(a) Pink shaped region is annotated and the blue square is tile image, eight yellow dots shows eight-points, which shown eight-points method will not extract this tile image. (b) Green contour indicating the annotated region, and the blue squares are the candidate tile image. But the red squares and red arrow indicating that the tiles images are not candidate for extraction. (c) Green contour also indicating the annotated region, blue squire indicating candidate tile images. But some candidate tile image are lots of blank area or fat shown in the shaded region.
Table 1.
Set up of IoT and 1 -BoT for tile images extraction of both training and testing set, number of tile images for both CAM and PAIP dataset.
Table 2.
Processing Time for CAM 111 positive slides and PAIP 50 slides, unit is in seconds.
Q1 is 25th percentile, Median is 50th percentile, Q3 is 75th percentile.
Fig 2.
Visualization for boundaries extraction.
Blur square are tile images we will extract. (a) IoT set at 1.0, square tile images are 100% within annotated regions, while there are some annotated region cannot be extracted. (b) IoT set at 0.1, greater than 10% intersecting area will be included. (c) IoT set at 0.2, which only very few difference between (b). (d) IoT set at 0.5, indicating inclusion tile image greater than 50% intersecting area. The red square is the tile not included compared wiht (b).
Fig 3.
Visualization exclusion blank tile image within annotated region.
Blue square tile images are extracted. (a) Showing in the red shaped area, blank tile images also included, b, c and d are zoom-in of the red shaped area. (b) BoT set at 0.1, showing that if the tile image is 90% blank area will excluded. (c) BoT set at 0.2, tile image with 80% blank area will be excluded, showing that more tile images exclude from a, b. (d) BoT set at 0.5, showing a lot more tile images at the red shaped area are not included.
Fig 4.
(a) Number of the positive tile images from set A to set H is decreasing. (b)Visualize performance model on validation set. (c) Visualize performance on hold-out testing set. (d) The relative difference in the performance between validation set and test set (z-score).
Table 3.
Pearson correlation coefficient for both validation set and testing set of CAM datasets.
Table 4.
Pearson correlation coefficient for both validation set and testing set of PAIP dataset.
Fig 5.
(a) Number of the positive tile images from set A to set H is decreasing. (b)Visualize performance model on validation set. (c) Visualize performance on hold-out testing set. (d) The relative difference in the performance between validation set and test set (z-score).