| --- |
| license: cc-by-nc-4.0 |
| tags: |
| - computational-pathology |
| - colorectal-liver-metastases |
| - tissue-segmentation |
| - whole-slide-imaging |
| library_name: keras |
| extra_gated_prompt: >- |
| These weights are derived from clinical research data and are released for |
| non-commercial research use only. They are not a medical device and must not |
| be used for clinical decision making. By requesting access you agree to these |
| terms. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| I agree to use these weights for non-commercial research only: checkbox |
| --- |
| |
| # CRLM liver tissue segmentation |
|
|
| Whole slide image (WSI) tissue segmentation for hematoxylin and eosin sections |
| of colorectal liver metastases (CRLM). This model produces the tissue mask used |
| upstream of the desmoplastic status and survival pipeline |
| (`computationalpathologygroup/pathology-liver-survival`). |
|
|
| ## Method |
| The model is a U-Net (Keras backend), depth 5, branching factor 5, with batch |
| normalization and valid padding. It is trained with categorical cross-entropy on |
| 412 by 412 patches at 0.5 micron pixel spacing, with RGB range normalization to |
| 0 to 1. Inference is run through the `applynetwork_multiproc.py` script from the |
| `pathology-fast-inference` code base. |
|
|
| Two networks run in sequence: |
| 1. `bg_network.net` (28 MB): background versus tissue segmentation, run first at |
| 2.0 micron spacing to create the tissue mask. |
| 2. `nn1_best_model.net` (87 MB): the main multi-class tissue segmentation, run at |
| 0.5 micron spacing using the tissue mask from step 1. |
|
|
| The network configuration in the code repository maps mask labels to 10 output |
| classes. The exact tissue class definitions should be confirmed with the authors |
| and against the Grand Challenge algorithm before interpretation. |
|
|
| ## Contents |
| - `nn1_best_model.net`: main multi-class tissue segmentation U-Net. |
| - `bg_network.net`: background versus tissue network used to create the tissue mask. |
|
|
| ## Intended use |
| Research use only. Not a medical device. Not for clinical decision making. |
|
|
| ## Links |
| - Code: https://github.com/DIAGNijmegen/pathology-liver-tissue-segmentation |
| - Grand Challenge: https://grand-challenge.org/algorithms/colorectal-liver-metastases-segmentation-in-he/ |
| - Downstream survival model: https://huggingface.co/computationalpathologygroup/pathology-liver-survival |
|
|
| ## Citation |
| Confirm the correct paper and citation with the authors before use. |
|
|
