Dataset Card for The Cancer Genome Atlas (TCGA) Multimodal Dataset

The Cancer Genome Atlas (TCGA) Multimodal Dataset is a comprehensive collection of clinical data, pathology reports, slide images, molecular data, and radiology images for cancer patients. This dataset aims to facilitate research in multimodal machine learning for oncology by providing embeddings generated using state-of-the-art models including GatorTron, MedGemma, Qwen, Llama, UNI, SeNMo, REMEDIS, and RadImageNet.

• Curated by: Lab Rasool
• Language(s) (NLP): English

Uses

from datasets import load_dataset

# Clinical data embeddings (4 models available)
clinical_gatortron = load_dataset("Lab-Rasool/TCGA", "clinical", split="gatortron")
clinical_medgemma = load_dataset("Lab-Rasool/TCGA", "clinical", split="medgemma")
clinical_qwen = load_dataset("Lab-Rasool/TCGA", "clinical", split="qwen")
clinical_llama = load_dataset("Lab-Rasool/TCGA", "clinical", split="llama")

# Pathology report embeddings (4 models available)
pathology_gatortron = load_dataset("Lab-Rasool/TCGA", "pathology_report", split="gatortron")
pathology_medgemma = load_dataset("Lab-Rasool/TCGA", "pathology_report", split="medgemma")
pathology_qwen = load_dataset("Lab-Rasool/TCGA", "pathology_report", split="qwen")
pathology_llama = load_dataset("Lab-Rasool/TCGA", "pathology_report", split="llama")

# Whole slide image embeddings
wsi_dataset = load_dataset("Lab-Rasool/TCGA", "wsi", split="uni")

# Molecular data embeddings
molecular_dataset = load_dataset("Lab-Rasool/TCGA", "molecular", split="senmo")

# Radiology embeddings (2 models available)
radiology_remedis = load_dataset("Lab-Rasool/TCGA", "radiology", split="remedis")
radiology_radimagenet = load_dataset("Lab-Rasool/TCGA", "radiology", split="radimagenet")

Example code for loading HF dataset into a PyTorch Dataloader. Note: Some embeddings are stored as buffers due to their multi-dimensional shape.

from datasets import load_dataset
import os
from torch.utils.data import Dataset
import numpy as np

class CustomDataset(Dataset):
    def __init__(self, hf_dataset):
        self.hf_dataset = hf_dataset

    def __len__(self):
        return len(self.hf_dataset)
    
    def __getitem__(self, idx):
        hf_item = self.hf_dataset[idx]
        embedding = np.frombuffer(hf_item["embedding"], dtype=np.float32)
        embedding_shape = hf_item["embedding_shape"]
        embedding = embedding.reshape(embedding_shape)
        return embedding

if __name__ == "__main__":
    
    # Load clinical embeddings from different models
    clinical_gatortron = load_dataset("Lab-Rasool/TCGA", "clinical", split="gatortron")
    clinical_llama = load_dataset("Lab-Rasool/TCGA", "clinical", split="llama")
    wsi_dataset = load_dataset("Lab-Rasool/TCGA", "wsi", split="uni")
    
    # Example: Access embeddings
    for index, item in enumerate(clinical_gatortron):
        embedding = np.frombuffer(item.get("embedding"), dtype=np.float32).reshape(item.get("embedding_shape"))
        print(f"GatorTron embedding shape: {embedding.shape}")  # Shape: (1024,)
        break
        
    for index, item in enumerate(clinical_llama):
        embedding = np.frombuffer(item.get("embedding"), dtype=np.float32).reshape(item.get("embedding_shape"))
        print(f"Llama embedding shape: {embedding.shape}")  # Shape: (2304,)
        break

Dataset Statistics

Clinical Data

• 10,771 patient records per model
• 113 columns including clinical metadata and embeddings
• Embedding dimensions:
  • GatorTron: 1024
  • MedGemma: 2560
  • Qwen: 1024
  • Llama: 2304
  • BioBERT: 768

Pathology Reports

• 10,857 patient records per model
• 17 columns including pathology metadata and embeddings
• Embedding dimensions:
  • GatorTron: 1024
  • MedGemma: 2560
  • Qwen: 1024
  • Llama: 2304

Dataset Creation

Data Collection and Processing

The raw data for this dataset was acquired using MINDS, a multimodal data aggregation tool developed by Lab Rasool. The collected data includes clinical information, pathology reports, and whole slide images from The Cancer Genome Atlas (TCGA). The embeddings were generated using the HoneyBee embedding processing tool, which utilizes foundational models such as GatorTron, MedGemma, Qwen, Llama, and UNI.

Who are the source data producers?

The source data for this dataset was originally collected and maintained by The Cancer Genome Atlas (TCGA) program, a landmark cancer genomics project jointly managed by the National Cancer Institute (NCI).

Citation

@article{honeybee,
      title={HoneyBee: A Scalable Modular Framework for Creating Multimodal Oncology Datasets with Foundational Embedding Models}, 
      author={Aakash Tripathi and Asim Waqas and Yasin Yilmaz and Ghulam Rasool},
      year={2024},
      eprint={2405.07460},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}
@article{waqas2024senmo,
  title={SeNMo: A self-normalizing deep learning model for enhanced multi-omics data analysis in oncology},
  author={Waqas, Asim and Tripathi, Aakash and Ahmed, Sabeen and Mukund, Ashwin and Farooq, Hamza and Schabath, Matthew B and Stewart, Paul and Naeini, Mia and Rasool, Ghulam},
  journal={arXiv preprint arXiv:2405.08226},
  year={2024}
}

For more information about the data acquisition and processing tools used in creating this dataset, please refer to the following resources:

• MINDS paper: https://pubmed.ncbi.nlm.nih.gov/38475170/
• MINDS codebase: https://github.com/lab-rasool/MINDS
• HoneyBee paper: https://arxiv.org/abs/2405.07460
• HoneyBee codebase: https://github.com/lab-rasool/HoneyBee/

Contact Information

For any questions or issues, please contact the dataset curators at [aakash.tripathi@moffitt.org].