There are three options for filtering tiles.
- 9tissues
- BRCATU
- TUNTINV
This network was trained on the publicly available NCT-CRC-HE-100K dataset, available here. It uses a 224px tile sampled at 0.5mpp and classifies it as one of 9 classes. It can use it to filter background tiles or other classes of interest by setting the following elements inside the create_tiles
in config file as:
remove_tiles: '9tissues'
remove_tiles_labels_exclude:
- BACK
Details regarding other detected tile classes can be found on its official Zenodo page.
This is a custom trained deep neural network ConvNext-tiny
, trained on a combination of publicly available data. In particular, examples of the TU
(Tumor class) were obtained and preprocessed in the form of tiles from annotations of TCGA slides, available here. Additionally, we use the BG
(Background) class of the the NCT-CRC-HE-100K dataset. The NT
class is generated by training on all tissues not recognized as one of those two classes.
It generates of the three classes for each slide:
- BG: Background tile
- NT: any other tissue not recognized as tumor
- TU: tissue from the tumor enviroment
remove_tiles: 'BRCATU'
remove_tiles_labels_exclude:
- BG
- NT
The TUNTINV filtering network is custom trained ConvNext-tiny
deep CNN, trained with a combination of the previously described BRCATU
Neural Network and manually annotated patches of invasive tumor. This network focuses on identifying tiles containing invasive tumor.
It also classifies tiles as the following class:
- BKG: background
- NOT_TUM: other tissues that are not tumor
- TUM_OTHER: tiles containing non-invasive tumor
- TUM_INV: tiles classified as invasive tumor
remove_tiles: 'TUNTINV'
remove_tiles_labels_exclude:
- BKG
- NOT_TUM
- TUM_OTHER