2025 Dr. Robert Gillies Machine Learning Workshop in Cancer - Hackathon Dataset

Dataset Description

This dataset is part of the 2025 Dr. Robert Gillies Machine Learning Workshop at Moffitt Cancer Center. It contains de-identified medical imaging, pathology reports, and clinical data for glioblastoma patients, designed for an overall survival prediction challenge.

Dataset Splits

• Training Set: 250 patients with complete survival outcomes
• Test Set: 50 patients with withheld survival outcomes

Training Dataset

• Total Patients: 250
• Pathology Images: 250 whole-slide tissue images (.tif format, ~512MB each)
• Clinical Records: 250 patient records with comprehensive clinical and pathology data
• Total Size: ~128GB

Test Dataset

• Total Patients: 50
• Pathology Images: 50 whole-slide tissue images (.tif format, ~280MB each)
• Clinical Records: 50 patient records with predictive features only (survival outcomes withheld)
• Total Size: ~14GB
• Purpose: Final evaluation with ground truth outcomes held by organizers

Dataset Structure

train/
├── train.csv                           # Clinical and pathology data (250 patients, 30 columns)
└── images/                             # Pathology images directory
    ├── P0002.tif                       # Patient pathology slide (~512MB)
    ├── P0005.tif
    ├── P0006.tif
    └── ... (250 total .tif files)

test/
├── test.csv                            # Clinical features only (50 patients, 21 columns)
└── images/                             # Pathology images directory
    ├── P0001.tif                       # Patient pathology slide (~280MB)
    ├── P0004.tif
    ├── P0007.tif
    └── ... (50 total .tif files)

Note: Images are whole-slide tissue scans in TIFF format. Each file is named with the corresponding patient ID (e.g., P0002.tif).

Clinical Data Schema (Training Set)

The train.csv file in the training set contains the following fields:

Demographics:

• patient_id: Unique de-identified patient identifier
• age_at_diagnosis: Patient age at diagnosis
• gender: Patient gender
• race: Patient race
• ethnicity: Patient ethnicity

Tumor Characteristics:

• primary_diagnosis: Primary cancer diagnosis
• tumor_grade: Tumor grade classification
• ajcc_m: AJCC M stage (metastasis)
• ajcc_n: AJCC N stage (lymph nodes)
• classification_of_tumor: Tumor classification (primary/recurrence/progression)
• morphology: ICD-O-3 morphology code
• tissue_origin: Anatomical site of origin
• laterality: Side of body affected
• prior_malignancy: History of prior cancer
• synchronous_malignancy: Concurrent cancer diagnosis

Clinical Outcomes:

• vital_status: Patient vital status (Alive/Dead)
• overall_survival_days: Overall survival time in days
• overall_survival_event: Survival event indicator (0/1)
• days_to_death: Time to death event
• days_to_last_followup: Time to last follow-up
• cause_of_death: Cause of death
• days_to_progression: Time to disease progression
• days_to_recurrence: Time to disease recurrence
• progression_or_recurrence: Indicator of disease progression/recurrence
• disease_response: Response to treatment

Treatment Information:

• treatment_types: Types of treatments received
• therapeutic_agents: Specific drugs/agents used
• days_to_treatment: Time to treatment initiation
• treatment_outcome: Outcome of treatment

Pathology:

• pathology_report: Complete de-identified pathology report text

Test Set Features

The test.csv file in the test set contains 21 predictive features only. The following 9 survival outcome columns have been removed to prevent participants from calculating C-index:

Withheld Columns:

• vital_status (Alive/Dead status)
• overall_survival_days (survival time - required for C-index)
• overall_survival_event (event indicator - required for C-index)
• days_to_death
• days_to_last_followup
• cause_of_death
• days_to_progression
• days_to_recurrence
• progression_or_recurrence

Available Test Features (21 columns):

• All demographic fields (patient_id, age_at_diagnosis, gender, race, ethnicity)
• All tumor characteristics (primary_diagnosis, tumor_grade, ajcc_m, ajcc_n, classification_of_tumor, morphology, tissue_origin, laterality, prior_malignancy, synchronous_malignancy)
• Treatment information (treatment_types, therapeutic_agents, days_to_treatment, treatment_outcome, disease_response)
• Pathology report (pathology_report)

Submission Format

For the overall survival prediction challenge, submit predictions as a CSV file with the following format:

patient_id,outcome,risk_score
P0001,1,0.742
P0004,0,0.234
P0007,1,0.891

Column Definitions:

• patient_id: Patient identifier (must match test set IDs)
• outcome: Predicted binary outcome (0 = censored/alive, 1 = event/death)
• risk_score: Predicted risk score (higher = higher risk of event)

Evaluation Metrics: C-Index (Concordance Index)

Usage

Quick Start

This is a multimodal dataset combining:

• Pathology images (TIFF format, individual whole-slide images)
• Clinical/tabular data (CSV format with patient metadata)

Storage Requirements:

• Full dataset: ~142GB
• Training split: ~128GB (250 images + CSV)
• Test split: ~14GB (50 images + CSV)

Loading Clinical Data

import pandas as pd

# Load clinical data directly
train_data = pd.read_csv("hf://datasets/Lab-Rasool/hackathon/train/train.csv")
test_data = pd.read_csv("hf://datasets/Lab-Rasool/hackathon/test/test.csv")

print(f"Training patients: {len(train_data)}")  # 250
print(f"Test patients: {len(test_data)}")      # 50

Downloading Images

Images can be downloaded individually or in bulk from the HuggingFace Hub:

from huggingface_hub import hf_hub_download
import os

# Download a specific patient's image
image_path = hf_hub_download(
    repo_id="Lab-Rasool/hackathon",
    filename="train/images/P0002.tif",
    repo_type="dataset",
    cache_dir="./hf_cache"
)

print(f"Image downloaded to: {image_path}")

# Or download all images in a split using snapshot_download
from huggingface_hub import snapshot_download

# Download entire train split (CSV + all images)
train_dir = snapshot_download(
    repo_id="Lab-Rasool/hackathon",
    repo_type="dataset",
    allow_patterns="train/*",
    cache_dir="./hf_cache"
)

print(f"Training data downloaded to: {train_dir}")

Loading Individual Images

import tifffile
from huggingface_hub import hf_hub_download

# Download and load a specific patient's image
patient_id = "P0002"
image_path = hf_hub_download(
    repo_id="Lab-Rasool/hackathon",
    filename=f"train/images/{patient_id}.tif",
    repo_type="dataset"
)

# Load the image
patient_image = tifffile.imread(image_path)
print(f"Image shape: {patient_image.shape}")
print(f"Image size: {patient_image.nbytes / (1024**2):.1f} MB")

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

• ✅ Academic and research use permitted
• ✅ Educational use permitted
• ❌ Commercial use prohibited
• ✅ Derivative works permitted with attribution

Important Notes

• Training Set (250 patients): Use for model development. Includes survival outcomes.
• Test Set (50 patients): Use for predictions. Survival outcomes are withheld.
• Images: Individual .tif files (~142GB total). Can be downloaded selectively or in bulk.
• Streaming: Images can be downloaded on-demand during training using HuggingFace's caching.
• Submissions: Must include all test patients with correct patient IDs.
• No Self-Evaluation: Test set ground truth is held by organizers only.