FLARE-MedFM/FLARE-Task5-MLLM-3D

FLARE Task 5: Multimodal Model for 3D Medical Image Parsing (Homepage)

The task is to train one multimodal model for report generation and vision QA.

Data Description

The dataset contains two subsets for abdomen and lung CT report generation and VQA.

FLARE-Task5-MLLM-3D/
├── README.md
├── train # training set
│   ├── CT-AMOS-1290 # source: https://era-ai-biomed.github.io/amos/
│   ├── CT-AMOS-Tr.json
│   ├── CT-RATE-2000 # source: https://huggingface.co/datasets/ibrahimhamamci/CT-RATE  we select unique patients with disease and clinical information.
│   └── CT-RATE-Tr.json
├── validation/ # validation sets 
│   ├── images
│   └── val.json
# To avoid potential data leakage, the testing set will not be released.

Other public datasets can be used but participants must submit an algorithm docker that only trained on the provided dataset. Each team can submit two dockers:

• only use challenge data: teamname.tar.gz
• use both challenge data and public data: teamname_publicdata.tar.gz

Dataset Download Instructions

Participants can download the complete dataset using the following Python script:

from huggingface_hub import snapshot_download

local_dir = "./FLARE-Task5-MLLM-3D"
snapshot_download(
    repo_id="FLARE-MedFM/FLARE-Task5-MLLM-3D",
    repo_type="dataset",
    local_dir=local_dir,
    local_dir_use_symlinks=False,
    resume_download=True,
)

TotalSegmentator Organ Segmentations

We also provide segmentations generated by TotalSegmentator for the AMOS-MM dataset. These can be downloaded here.

Reference

• Ji, Yuanfeng, et al. "Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation." Advances in Neural Information Processing Systems 35 (2022): 36722-36732.
• Hamamci, Ibrahim Ethem, et al. "Developing generalist foundation models from a multimodal dataset for 3D computed tomography." arXiv preprint arXiv:2403.17834 (2024).
• Hamamci, Ibrahim Ethem, et al. "Generatect: Text-conditional generation of 3d chest ct volumes." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.
• Hamamci, Ibrahim Ethem, Sezgin Er, and Bjoern Menze. "Ct2rep: Automated radiology report generation for 3D medical imaging." International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024.