PathoSlot

Official implementation of PathoSlot, accepted by MICCAI 2026.

Installation

Create a Python environment and install the required packages:

conda create -n pathoslot python=3.10
conda activate pathoslot

pip install torch numpy pandas h5py pyyaml tqdm matplotlib seaborn scikit-learn scikit-survival

Install the CUDA-enabled PyTorch build that matches your system if you plan to train on GPU.

Data Format

The training scripts expect:

• A split CSV/TSV with sample IDs, fold columns, biomarker labels, and survival labels.
• A task config YAML. See my_datatsets/zhongshan_config.yaml for the expected fields.
• One .h5 embedding file per sample ID in the embedding directory.

Each .h5 file should contain:

• vis_features
• text_features
• name_embedding

Run

Set paths with environment variables:

export PATHOSLOT_SPLIT=/path/to/split.csv
export PATHOSLOT_TASK_CONFIG=/path/to/config.yaml
export PATHOSLOT_EMBEDDINGS_DIR=/path/to/embeddings
export PATHOSLOT_RESULTS_DIR=/path/to/output_dir

Train and evaluate:

python train_slot.py

Evaluate an existing checkpoint:

python test_slot.py

By default, the scripts fall back to repository-relative placeholder paths under data/ and outputs/.