Tissue-level Prediction

The pipeline run_tissue_level_tasks tests a model’s ability to capture clinically relevant local and global tissue features with pathology tasks such as cancer grading, subtyping, and treatment response prediction. The model embeds each tissue as a sequence of tile embeddings; a gated attention-based multiple-instance learning (ABMIL) classifier trained on the training patients predicts the patient-level label of the test patients.

Running the Benchmark

python -m spora_bench.benchmarks.run_tissue_level_tasks \
  model_config=<path-to-model-config> \
  datasets_config=<path-to-datasets-config> \
  benchmarks_config=<path-to-benchmarks-config>

For example, all tissue-level tasks on cords2024cancer using VirTues:

python -m spora_bench.benchmarks.run_tissue_level_tasks \
  model_config=configs/models/virtues.yaml \
  datasets_config=configs/datasets/virtues/cords2024cancer.yaml \
  benchmarks_config=configs/benchmarks/tissue_level/cords2024cancer.yaml

As with cell phenotyping, batched submission of dataset and benchmark configs is possible via lists or wildcards.

The ABMIL architecture and training settings (learning rate, epochs, early stopping, validation fraction) are read from configs/downstream_models/abmil.yaml.

Outputs

Tile embeddings are cached in output_dir/<model_name>/tile_tokens/. For each task, a classification report and a bootstrapped classification report (1,000 resamples) are saved to output_dir/<model_name>/results/.

Benchmark Config

datasets:
  <dataset_name>:
    benchmarks:
      tissue_level:
        <task_name>:
          label_col: <column_name>
          excluded_classes: <list of classes to exclude>

label_col is a column of metadata/tissues.parquet. Classes listed in excluded_classes are filtered out prior to training and evaluation.

The datasets and configs used in the paper are listed in the Task Zoo.