khoaguin/chest-ct-segmentation

Chest CT Segmentation - Federated Learning Partitions

Dataset partitions for Federated SAM2-LoRA medical image segmentation across multiple Data Owners.

Dataset Description

This dataset contains partitioned chest CT scans designed for federated learning experiments with heterogeneous clients. Each partition represents a different hospital/data owner with varying data availability and training capabilities.

Source

Original data from Chest CT Segmentation on Kaggle.

Federated Learning Setup

Data Owner	Type	Training Method	Contributes to FedAvg
DO1	Zero-shot	CLIP text prompts	No
DO2	Few-shot	Memory bank	No
DO3	LoRA	Gradient training	Yes
DO4	LoRA	Gradient training	Yes

Dataset Structure

├── do2_fewshot/
│   ├── mock/          # Sample data for testing
│   │   ├── train/
│   │   └── test/
│   ├── private/       # Full training data
│   │   ├── train/
│   │   └── test/
│   └── README.md
├── do3_lora/
│   ├── mock/
│   ├── private/
│   └── README.md
├── do4_lora/
│   ├── mock/
│   ├── private/
│   └── README.md
└── README.md

Each split contains:

• images/ - RGB JPEG chest CT slices
• masks/ - Binary segmentation masks
• train.csv or test.csv - Image-mask mapping

Statistics

Partition	Private (train/test)	Mock (train/test)	Unique Patients
do2_fewshot	4 / 2	2 / 1	6
do3_lora	30 / 8	8 / 2	38
do4_lora	28 / 8	7 / 2	36
Total	62 / 18	17 / 5	80

• Diversity: 1 image per patient (no overlap between DOs)
• Mock data: Half of private data for testing/simulation

Usage

Load with Hugging Face

from huggingface_hub import snapshot_download

# Download all partitions
snapshot_download(
    repo_id="khoaguin/chest-ct-segmentation",
    repo_type="dataset",
    local_dir="./dataset"
)

# Download specific partition
snapshot_download(
    repo_id="khoaguin/chest-ct-segmentation",
    repo_type="dataset",
    allow_patterns="do3_lora/**",
    local_dir="./dataset"
)

# Download mock data only
snapshot_download(
    repo_id="khoaguin/chest-ct-segmentation",
    repo_type="dataset",
    allow_patterns="*/mock/**",
    local_dir="./dataset"
)

Load Images

import pandas as pd
from PIL import Image
from pathlib import Path

# Load DO3 training data
data_path = Path("dataset/do3_lora/private/train")
df = pd.read_csv(data_path / "train.csv")

for _, row in df.iterrows():
    image = Image.open(data_path / "images" / row["ImageId"])
    mask = Image.open(data_path / "masks" / row["MaskId"])

Data Format

CSV Structure

ImageId,MaskId
ID00131637202220424084844_30.jpg,ID00131637202220424084844_mask_30.jpg

Image Details

• Format: JPEG
• Type: Grayscale CT slices (stored as RGB)
• Resolution: Variable (typically 512x512)

Mask Details

• Format: JPEG
• Type: Binary segmentation mask
• Values: 0 (background), 255 (lung/tissue)

Citation

If you use this dataset, please cite:

@misc{chest-ct-segmentation-fl,
  title={Chest CT Segmentation for Federated Learning},
  author={OpenMined},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/khoaguin/chest-ct-segmentation}
}

Related Resources

• FL-SAM2-LoRA Project
• Original Kaggle Dataset
• SAM2 Paper

License

This dataset is released under CC BY 4.0.