{
  "repository": "https://github.com/digitalpathologybern/hover_next_train",
  "revision": "db4adfa425ffcad51d01a361e660a40c1ecda9dd",
  "classes": {
    "0": "background",
    "1": "neutrophil",
    "2": "epithelial-cell",
    "3": "lymphocyte",
    "4": "plasma-cell",
    "5": "eosinophil",
    "6": "connective-tissue-cell",
    "7": "mitosis"
  },
  "evidence": {
    "src/constants.py": "CLASS_NAMES defines the seven foreground names in order.",
    "src/data_utils.py": "add_3c_gt_fast reads Y[...,0] as instances; get_weighted_sampler reads gt[...,1] as classes 0..7. get_lizard reads source fold_0 train/valid and separate test arrays.",
    "src/post_proc_utils.py": "convert_reg counts object classes == i+1 for class_names[i], establishing one-based foreground mapping."
  },
  "source_array_scope": "Only lizard_mitosis.zip, not separate mitosis_ds/eos_val/mit_val archives. Source array index is the only per-image identifier available in this ZIP."
}