Contribute

You can support the project by adding graph models or increasing the number of features used. Every contribution must have type hints and proper docstrings.

Expanding the graph module

Models

In case you are interested in adding more models to the graph zoo, you can do so modifying the source code and creating a pull request. Models should go in the folder tumourkit/classification/models and must adhere to the following API:

class NewGraph(nn.Module):
  def __init__(self, in_feats, h_feats, num_classes, num_layers, drop_rate, norm_type):
      super(NewGraph, self).__init__()
      # Add any layers you want here
      # ...
      # Dropout layers should be defined as
      # nn.Dropout(drop_rate)
      # Normalization layers should be defined a
      # Norm(norm_type=norm_type, hidden_dim=h_feats)

  def forward(self, g, in_feat):
      h = in_feat
      # Make any computation with the hidden embedding h here
      # ...
      return h

Your model should have a variable amount of layers controlled by num_layers, a variable amount of dropout rate as given by drop_rate and must include either batch normalization or no normalization, both controlled through Norm.

Extracted features

Currently, the features that are used by the models are:

The area and perimeter of the cell, in pixels.
The standard deviation of the pixel values in gray format.
The histogram of the red, green and blue channels quantized into five bins each.
The prior probability of the class as given by Hovernet.

If you want to add more features, you should change two files: tumourkit/classification/train_graphs.py, tumourkit/utils/preprocessing.py. In the first one you need to modify the load_model function.

def load_model(conf: Dict[str,Any], num_classes: int) -> nn.Module:
    # ...
    num_feats = 18 + (1 if num_classes == 2 else num_classes)
    # ...

In there modify the num_feats variable to denote the number of features there are. If you add 2 more features, change the 18 by a 20.

The other function needed to be adapted is extract_features:

def extract_features(msk_img: np.ndarray, bin_msk: np.ndarray, debug=False) -> Dict[str, np.ndarray]:
    # ...
    feats = {}
    # Add any feature you want to the dictionary feats.
    # You can use the keys you want, they will be reflected as columns in the saved files.
    return feats