Machine Learning models for a fast.ai v1 ResNet18 U-Net CNN, to be used with the HistoFlow tool. See the to-be-released thesis and paper for details.
Copy the models to the HistoFlow server from the pathology-ml-model-training repository. While it is not yet possible to choose external models in the UI (as of May 27th 2020), the feature is planned and meanwhile the model can be hardcoded to be the default model in the server (in server/main.py).
For each model trained_model.pth and input/export.pkl is provided. The first is used for easy inference and the second is provided to be able to use the model for transfer learning and retrain it.
a0e6aaa83fb7a50ab5de37faef9fecb7-557c183ee44cafc2bf48a20e24543710 is our best performing model for cell instance segmentation on breast cancer fluorescence images. It was first trained with artificial data and then fine tuned with manual annotations from real images. This model is reffered to as Model D in the thesis and the paper.
7c0c7084ac8007ab0c24a3ee563e349c-d1902cca8d8c72222dc5315c1411a337-92f2cf69f5abe9ea11c31c03f6d8cb23 is based on the best performing instance segmentation model and was extended with manual annotations to also do classification of each cell into epithelial or not. The classification result is returned in the third channel per pixel. This is the model used for the classification evaluation in the thesis and paper.