CCDI-MCI Whole-Slide Image Embeddings

Overview

This release contains pretrained feature embeddings generated from 4,570 whole-slide images in the CCDI Molecular Characterization Initiative (CCDI-MCI) collection.

The embeddings are intended to support research workflows such as cohort exploration, retrieval, and classification. They are not clinical predictions and have not been validated for diagnostic use.

Data

Data used in this study was obtained from the National Cancer Institute's Childhood Cancer Data Initiative (CCDI) Molecular Characterization Initiative (MCI) [2]. Whole-slide images (WSIs) were obtained from the CCDI-MCI collection hosted by the NCI Imaging Data Commons (IDC) [1,3]. The collection is also identified by doi:10.5281/zenodo.11099086 [4].

DICOM Slide Microscopy series were downloaded on December 15, 2025, using the idc-index CLI client [5] from IDC data release v23. Each selected series was downloaded using its crdc_series_uuid

Downloads and subsequent feature extraction was done on the NIH HPC Biowulf cluster. The download included 4,576 SM series from 4,407 patients, totaling 21,363 DICOM instances across the associated images and pyramid levels. Six series were subsequently excluded because QC steps filtered all available patches, leaving 4,570 WSIs in this specific release.

Models

The initial release includes embeddings from three Mahmood Lab models:

• conch_v15 [6]
• uni2h [7]
• titan [8]

CONCH v1.5 [6,8] and UNI2-h [7] files contain patch-level embeddings and coordinates. TITAN [8] files contain slide-level embeddings together with the coordinates used during patch feature extraction.

Repository layout

Files are organized by model and patient.

embeddings/
  {model}/
    {patient_id}/
      {SeriesInstanceUID}.h5

The DICOM SeriesInstanceUID is the primary link back to the corresponding raw image in the IDC. The manifest csv maps each embedding to its patient, study, series, and source DICOM file.

Minimal retrieval example

Download all the patients for a given model either with the hf CLI or python API. Patient level retrevial is demonstrated on our GitHub

hf download CBIIT-CGBB/CCDI-MCI \
  --repo-type dataset \
  --include "embeddings/<model>/*/*.h5" \
  --local-dir /path/to/output/dir

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id='CBIIT-CGBB/CCDI-MCI',
    repo_type="dataset",
    allow_patterns=[
        f"embeddings/{model}/*/*.h5",
        "metadata/slide_manifest.csv", # include the manifest as well
    ],
    local_dir=/path/to/output/dir,
)

Obtaining Raw Images

The original raw WSI DICOM pyramids can be obtained with the idc-index tool and either of the following commands:

idc download `{SeriesInstanceUID}`

idc download {PatientID}

Processing and Feature Extraction

Image processing and feature extraction was performed with the TRIDENT toolkit [9]. This includes tissue detection and artifact filtering performed with the GrandQC tissue-detection and multi-class artifact-segmentation models [10] GeoJSONs from the resultant segmentations are available in segmentations/{patient_id}/{SeriesInstanceUID}.geojson

HDF5 contents

Each HDF5 file contains:

• features: patch-level or slide-level embeddings
• coords: level-0 pixel coordinates associated with the extracted patches
• metadata describing the source image, release, extraction pipeline, model weights, and embedding type

The exact feature shape depends on the model. Users should inspect the dataset shape before use.

Intended use

including but not limited to:

• Research on computational pathology methods
• Training or evaluating downstream classificaiton models
• Patient or series-based embedding retrieval
• Linking embeddings with clinical or molecular data

Some example jupyter noteboooks demostrating embedding retrevial and downstream use can be found on our assoicated GitHub repo

Associated preprint

A full preprint describing the construction and validation of this release is in preparation and has not yet been published.

References

1. Fedorov A, Longabaugh WJR, Pot D, et al. NCI Imaging Data Commons. Cancer Research. 2021;81(16):4188-4193. doi:10.1158/0008-5472.CAN-21-0950.
2. National Cancer Institute. The Childhood Cancer Data Initiative Molecular Characterization Initiative.
3. NCI Imaging Data Commons. CCDI-MCI collection.
4. NCI Imaging Data Commons. CCDI-MCI original data. Zenodo. doi:10.5281/zenodo.11099086.
5. Imaging Data Commons. idc-index: programmatic access to NCI Imaging Data Commons.
6. Lu MY, Chen B, Williamson DFK, et al. A visual-language foundation model for computational pathology. Nature Medicine. 2024;30(3):863-874. doi:10.1038/s41591-024-02856-4.
7. Chen RJ, Ding T, Lu MY, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine. 2024;30(3):850-862. doi:10.1038/s41591-024-02857-3.
8. Ding T, Wagner SJ, Song AH, Chen RJ, et al. A multimodal whole-slide foundation model for pathology. Nature Medicine. 2025;31(11):3749-3761. doi:10.1038/s41591-025-03982-3.
9. Zhang A, Jaume G, Vaidya A, Ding T, Mahmood F. Accelerating Data Processing and Benchmarking of AI Models for Pathology. arXiv. 2025. arXiv:2502.06750. doi:10.48550/arXiv.2502.06750.
10. Weng Z, Seper A, Pryalukhin A, et al. GrandQC: A comprehensive solution to quality control problem in digital pathology. Nature Communications. 2024;15:10685. doi:10.1038/s41467-024-54769-y.