README.md

---
language:
- en
license: other
pretty_name: TC-SSA WSI Feature Bags
size_categories:
- 1K<n<10K
task_categories:
- image-feature-extraction
tags:
- computational-pathology
- whole-slide-imaging
- multiple-instance-learning
- conch
- cpathpatchfeature
configs:
- config_name: file-index
  data_files:
  - split: train
    path: metadata/file_index.csv
---
# TC-SSA: Token Compression via Semantic Slot Aggregation for Gigapixel Pathology Reasoning

[![Python 3.8+](https://img.shields.io/badge/python-3.8+-blue.svg)](https://www.python.org/downloads/)
[![PyTorch](https://img.shields.io/badge/PyTorch-2.0+-red.svg)](https://pytorch.org/)
[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
[![Project Page](https://img.shields.io/badge/Project-Homepage-007c89.svg)](https://ozzychen97.github.io/TC-SSA/)
[![arXiv](https://img.shields.io/badge/arXiv-2603.01143-b31b1b.svg)](https://arxiv.org/pdf/2603.01143)

**Links:** [Project homepage](https://ozzychen97.github.io/TC-SSA/) | [arXiv paper](https://arxiv.org/pdf/2603.01143) | [Code](https://github.com/OzzyChen97/TC-SSA)

**Authors:** [Zhuo Chen](https://orcid.org/0009-0000-6089-1368)<sup>1,2</sup>, [Xiaoyu Yang](https://orcid.org/0000-0003-0273-9573)<sup>1</sup>, and [Lijian Xu](https://orcid.org/0000-0002-6632-4011)<sup>1,*</sup>

<sup>1</sup> Shenzhen University of Advanced Technology, Shenzhen, Guangdong, China  
<sup>2</sup> University of Nottingham Ningbo China, FoSE, Ningbo, Zhejiang, China  
<sup>*</sup> Corresponding author: [xulijian@suat-sz.edu.cn](mailto:xulijian@suat-sz.edu.cn)

---

# TC-SSA WSI Feature Bags

This public repository contains pre-extracted whole-slide image patch features
used by TC-SSA and SSAMIL experiments. Binary feature bags are organized by
feature source and encoder. The Dataset Viewer intentionally displays the file
index in `metadata/file_index.csv`; it does not attempt to interpret the HDF5
or PyTorch feature files as tabular datasets.

## Inventory

### CONCHv1.5 features

| Dataset | Files | Magnification / patch size | Format | Feature dim |
| --- | ---: | --- | --- | ---: |
| BRACS | 547 | 20x / 256 px | HDF5 | 768 |
| CAMELYON17 | 1,000 | 40x / 256 px | HDF5 | 768 |
| TCGA-BRCA | 1,129 | 20x / 256 px | HDF5 | 768 |
| TCGA-LUAD | 540 | 20x / 256 px | HDF5 | 768 |

Each HDF5 file contains a `features` dataset with shape
`(num_patches, 768)`. Some files also contain patch coordinates.

### CPathPatchFeature TCGA-BRCA subset

| Encoder / artifact | Files | Format | Feature dim |
| --- | ---: | --- | ---: |
| CHIEF / CTransPath | 1,133 | PyTorch tensor | 768 |
| GigaPath | 1,133 | PyTorch tensor | 1,536 |
| ResNet-50 | 1,133 | PyTorch tensor | 1,024 |
| UNI v1 | 1,133 | PyTorch tensor | 1,024 |
| Patch coordinates | 1,133 | HDF5 | n/a |

The four encoder directories have identical slide filename sets. The 1,125
slides used by the formal TCGA-BRCA experiment splits were validated for full
coverage and expected feature dimensions. Missing files from the upstream BRCA
subset were locally re-extracted and merged with the released collection.

## Layout

```text
features/
  conch_v15/
    BRACS/*.h5
    CAMELYON17/*.h5
    TCGA-BRCA/*.h5
    TCGA-LUAD/*.h5
  cpath_patch_feature/
    brca/
      chief/pt_files/*.pt
      gigap/pt_files/*.pt
      r50/pt_files/*.pt
      uni/pt_files/*.pt
      patches/*.h5
metadata/
  file_index.csv
```

## Dataset Viewer and access

The Viewer exposes a searchable index rather than loading hundreds of
gigabytes of heterogeneous binary tensors:

```python
from datasets import load_dataset

index = load_dataset("OzzyChen97/TC-SSA", "file-index", split="train")
print(index[0])
```

Download selected files with `hf download`, then load HDF5 features with
`h5py.File(...)` or PyTorch tensors with `torch.load(..., map_location="cpu")`.

## Provenance and references

The CPath encoder features and patch coordinates are derived from
[Dearcat/CPathPatchFeature](https://huggingface.co/datasets/Dearcat/CPathPatchFeature),
released with the work
[Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology](https://arxiv.org/abs/2506.02408).
Please cite that repository and its associated publications when using the
CPath subset.

The CONCH feature bags use
[CONCH](https://www.nature.com/articles/s41591-024-02856-4), a pathology
vision-language foundation model. The CPath collection uses
[UNI](https://www.nature.com/articles/s41591-024-02857-3),
[CHIEF](https://www.nature.com/articles/s41586-024-07894-z),
[GigaPath](https://www.nature.com/articles/s41586-024-07441-w), and
[ResNet-50](https://arxiv.org/abs/1512.03385).

Original WSI datasets and access pages:

- [BRACS dataset and paper](https://www.bracs.icar.cnr.it/background/)
- [CAMELYON17 challenge data](https://camelyon17.grand-challenge.org/Data/)
- [TCGA-BRCA at the NCI Genomic Data Commons](https://portal.gdc.cancer.gov/projects/TCGA-BRCA)
- [TCGA-LUAD at the NCI Genomic Data Commons](https://portal.gdc.cancer.gov/projects/TCGA-LUAD)

The upstream CPathPatchFeature repository additionally contains derived
features for TCGA-BLCA, TCGA-NSCLC, CPTAC-NSCLC, CAMELYON, and PANDA. Those
additional cohorts are referenced upstream and are not duplicated in this
TC-SSA repository.

## Citation

```bibtex
@inproceedings{chen2026tcssa,
  title     = {TC-SSA: Token Compression via Semantic Slot Aggregation for Gigapixel Pathology Reasoning},
  author    = {Chen, Zhuo and Yang, Xiaoyu and Xu, Lijian},
  booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
  publisher = {Springer Nature},
  series    = {Lecture Notes in Computer Science},
  year      = {2026},
  doi       = {10.48550/arXiv.2603.01143},
  url       = {https://arxiv.org/abs/2603.01143}
}
```

Please also cite the original WSI dataset and encoder papers relevant to the
files used in your study.