Artificial Intelligence and Machine Learning for Digital Pathology by Unknown

Author:Unknown
Language: eng
Format: epub
ISBN: 9783030504021
Publisher: Springer International Publishing

The first notebook data-handling-usage-guide comes with a small library, written by us, to easily access WSIs programmatically. It can be used as-is, or serve as an example on how to use the openslide interface in combination with the XML files for positive slides to find regions with tumorous tissue. It is mostly intended to give students a feeling of how to process WSIs. WSIs typically have a size of approximately 200.000 100.000 pixels. A practical approach to find out which regions contain tumorous tissue is to slice the whole WSI into smaller tiles. We will refer to this process as tiling. After that, we can train a CNN to predict whether a tile contains metastases (positive) or not (negative). Combining predictions of all tiles of a WSI will give us a heatmap. Figure 2 shows a WSI of the Camelyon16 data set labeled containing tumorous tissue. Executing the example code snippets of the first notebook we can, for example, get a thumbnail of the WSI (top left), a tumor mask, created with the information from the XML file (top right) and tiles of the regions containing tumorous tissue with varying overlapping parameter (bottom). So from the figure we can already see, on which things the notebook gives food for thought concerning the tiling process: Which size shall the tiles have? Which magnification level to chose? Shall they overlap or not, if so how much? For positive slides, what minimum percentage of the pixels must be tumorous so we label them as positive for our training set? For negative slides, how much percentage of the pixels must contain tissue (in contrast to slide background) so we use them as negative training tiles?

Because of the size of the data set, we will have to read in batches of tiles from hard drive (opposed to RAM for small toy data sets) when training the classifier. Additionally, finding suitable tiles is a computationally demanding task as we have to filter out tiles containing mostly background for negative WSIs and tiles containing not enough tumorous tissue for positive WSIs. As slides from different institutions or even different staining processes can vary a lot, e.g. in color, we also would prefer to compose a batch of tiles of as many different WSIs as possible to get a good representation of the whole data set in each batch. To fulfill these requirements while keeping the time needed to fetch batches of tiles as low as possible, we propose to create a custom training set of tiles beforehand.

This is done in the notebook create-custom-dataset. The notebook gives two options: Option A uses the highest possible magnification level (level 0) and resulting tiles will be of size 312 312 pixels, whereas option B only uses magnification level 2 and resulting tiles will only have a size of 256 256 pixels. The reason for the different sizes is, that we want positive slides to contain at least 60% tumorous tissue. If working on magnification level 3, the number of tiles containing 60% tumorous tissue will be a lot less the bigger the tile size.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.