shannonhill/HyperMIL-data

HyperMIL Reproducibility Data

Frozen inputs for HyperMIL: Adaptive Spatial–Semantic Hypergraph MIL for Survival Prediction from Whole-Slide Histopathology.

Code, pretrained models, patient splits, and evaluation instructions

This repository provides the input files used by the shared evaluation framework to reproduce the main comparison and ablation tables. It contains 12,096 files, totaling 142,866,662,028 bytes (142.87 GB / 133.05 GiB), across TCGA-STAD, TCGA-THCA, TCGA-CHOL, and TCGA-LIHC (1,344 WSIs).

Contents

data/
  WSI_data/
    TCGA-STAD/                # also TCGA-THCA, TCGA-CHOL, TCGA-LIHC
      patch_ft/*.npy          # frozen 512-D float32 patch features
      patch_coor/*.npy        # aligned patch coordinates
  hypergraphs/
    STAD_MGAS_V2/*.npz        # spatial incidence; all four cohorts
    STAD_PASG_V2_FINAL/*.npz   # semantic incidence; all four cohorts
  patchgcn_graphs/
    TCGA-STAD/*.npz           # lossless connectivity; all four cohorts
  KNN/
    TCGA-STAD/                # all four cohorts
      spatial/*.npz
      feature/*.npz
  kmeans_groups/
    Spatial-KMeans-GroupPool/cache/STAD/*.npz
    Semantic-KMeans-GroupPool/cache/STAD/*.npz
                             # also THCA, CHOL, LIHC
README.md
data_manifest.csv
dataset_info.json

Patch features are shared by all methods and stored only once. The compact PatchGCN files contain connectivity, node counts, and a reference feature-file hash; its model reads node features from WSI_data. Compaction preserved all 20 PatchGCN fold C-indices and the four-decimal paper table results.

Reproduction

Clone the code repository and follow its installation instructions. Its data downloader installs these files directly into the relative data/ directories expected by configs/paths.yaml. No manual path changes are needed with the supplied layout. The code release pins a dataset commit so subsequent repository changes do not silently change evaluation inputs.

The patient-level five-fold splits and fixed patch selections are distributed with the code. This dataset contains frozen inputs, not raw whole-slide images or preprocessing source code. Model weights are distributed in the GitHub release.

Use the same feature files, patch ordering, graph inputs, splits, and YAML configurations as the released evaluation. TCGA is the source of the underlying cohort data; refer to the NCI Genomic Data Commons for source data and applicable access policies. This repository does not grant additional rights over the underlying source data.

data_manifest.csv lists each released file's relative path, size, and SHA256. dataset_info.json records dataset totals and the shared patient split identity.