tumor-vqa/DeepTumorVQA_1.0

🧠 Overview

We present DeepTumorVQA, a diagnostic visual question answering (VQA) benchmark targeting abdominal tumors in CT scans. It comprises 9,262 CT volumes (3.7M slices) from 17 public datasets, with 395K expert-level questions spanning four categories: Recognition, Measurement, Visual Reasoning, and Medical Reasoning.

---

🧾 Dataset CT Volumes Overview

The following public abdominal CT datasets are included in DeepTumorVQA.
Note: The number of volumes may differ from the original publications due to validation splits or removal of duplicates.

Dataset (Year) [Source]	# of Volumes	# of Centers	Dataset (Year) [Source]	# of Volumes	# of Centers
1. CHAOS (2018) 🔗	20	1	2. Pancreas-CT (2015) 🔗	42	1
3. BTCV (2015) 🔗	47	1	4. LiTS (2019) 🔗	131	7
5. CT-ORG (2020) 🔗	140	8	6. WORD (2021) 🔗	120	1
7. AMOS22 (2022) 🔗	200	2	8. KiTS (2020) 🔗	489	1
9–14. MSD CT Tasks (2021) 🔗	945	1	15. AbdomenCT-1K (2021) 🔗	1,050	12
16. FLARE’23 (2022) 🔗	4,100	30	17. Trauma Detect. (2023) 🔗	4,711	23

To facilitate alignment between our VQA dataset and the original CT image sources, we follow the AbdomenAtlas naming rule and provide a mapping file that links each image ID in our dataset to its corresponding source identifier.

You can view the ID mapping CSV here: AbdomenAtlas_ID_mapping.csv

This file ensures traceability and reproducibility when working with external data references and annotations.

You may also email zzhou82@jh.edu for mapped full data and opportunities to collaborate in our future publications!

📁 Dataset QA Format

Each example in Tumor_VQA_dataset_V3.csv contains the following fields:

• question_id: A unique integer identifier for each VQA sample (e.g., 0).
• image_id: A string identifier for the corresponding CT volume or slice (e.g., BDMAP_00000001).
• spacing: Image voxel spacing (e.g., "[0.8222656 0.8222656 2.5 ]"), stored as a string.
• shape: The image dimensions (e.g., "(512, 512, 339)"), stored as a string.
• sex: Binary patient sex (Male, Female).
• age: Patient age in years, stored as a float64 (e.g., 65.0).
• scanner: Type of CT scanner used (e.g., siemens.
• contrast: Indicates use of contrast agent (Non-contrast, Arterial, Venous, etc.).
• question: A natural-language question about the image.
• answer: The corresponding expert-level answer to the question.
• multiple_choice_question: Reformulation of the question as a multiple-choice item.
• correct_option: The correct answer among multiple choices (a value from A to D).
• organ: The anatomical structure referenced in the question.
• lesion: The type of lesion involved (tumor, cyst, etc.).
• question_type: The general category of the question (recognition, measurement, visual reasoning, medical reasoning, etc.).
• question_subtype: A more granular subclassification (e.g., lesion_counting, organ_hu_measurement, `lesion_type_classification, etc.).
• split: Designates whether the sample belongs to the train or validation set.

---

🔍 Installation

You can load **DeepTumorVQA 1.0** directly using the 🤗 `datasets` library:

pip install datasets

from datasets import load_dataset

deep_tumor_vqa = load_dataset("tumor-vqa/DeepTumorVQA_1.0")
print(deep_tumor_vqa)

Acknowledgement and Disclosure of Funding

This work was supported by the Lustgarten Foundation for Pancreatic Cancer Research and the Patrick J. McGovern Foundation Award.

---

Citation

@article{chen2025vision,
  title={Are Vision Language Models Ready for Clinical Diagnosis? A 3D Medical Benchmark for Tumor-centric Visual Question Answering},
  author={Chen, Yixiong and Xiao, Wenjie and Bassi, Pedro RAS and Zhou, Xinze and Er, Sezgin and Hamamci, Ibrahim Ethem and Zhou, Zongwei and Yuille, Alan},
  journal={arXiv preprint arXiv:2505.18915},
  year={2025}
}

---