In this repository we provide code and example notebooks to assist participants start their algorithm development for the CoNIC challenge. In particular we provide:

• Evaluation code

  • Segmentation & classification: multi-class panoptic quality (mPQ⁺)
  • Predicting cellular composition: multi-class coefficient of determination (R²)

• Example notebooks

  • Data reading and simple dataset statistics
  • HoVer-Net baseline inference

NEWS: We have now released the training code that we used to train the baseline method (HoVer-Net). For this, we created a new branch, named conic in the original HoVer-Net repository. Click on this link to access the code!

To appropriately calculate the metrics, ensure that your output is in the following format:

• Instance Segmentation and classification map:

  • .npy array of size Nx256x256x2, where N is the number of processed patches.
  • First channel is the instance segmentation map containing values ranging from 0 (background) to n (number of nuclei).
  • Second channel is the classification map containing values ranging from 0 (background) to 6 (number of classes in the dataset).

• Composition prediction:

  • Single .csv file where the column headers should be:
    • neutrophil
    • epithelial
    • lymphocyte
    • plasma
    • eosinophil
    • connective
  • To make sure the calculation is done correctly, ensure that the row ordering is the same for both the ground truth and prediction csv files.

To get the stats for segmentation and classification, run:

python compute_stats.py --mode="seg_class" --pred=<path_to_results> --true=<path_to_ground_truth>

To get the stats for cellular composition prediction, run:

python compute_stats.py --mode="regression" --pred=<path_to_results> --true=<path_to_ground_truth>