BTUMQA: Brain Tumor MRI Visual Question Answering Dataset

Official dataset repository for the paper: "AUGR-VQA: Question-Adaptive Uncertainty-Guided Regional Reasoning for Reliable Brain Tumor MRI Visual Question Answering".

BTUMQA is a large-scale visual question answering dataset designed for controlled research evaluation of brain tumor MRI scans. It contains 225,000 QA pairs mapped to 2D MRI slices. The patient slice information and base images are derived directly from the RSNA BraTS 2021 2D dataset on Kaggle.

🔗 Important Links

• Paper Code Repository (GitHub): shuvomonowar00/AUGR-VQA (Private, publishing soon)
• Source Image Dataset (Kaggle): RSNA BraTS 2021 2D

This dataset is a controlled research evaluation resource and is not intended for clinical diagnosis, patient treatment, or deployment in real clinical workflows.

---

📂 Dataset Structure

The dataset files are organized into the btumqa_225k/ subfolder on the main branch to support versioning and benchmarking:

• btumqa_225k/btumqa_225k_qa_pairs.csv: The complete pool of 225,000 question-answer pairs with image references.
• btumqa_225k/btumqa_225k_train.csv: The training split used in the AUGR-VQA study (157,501 QA pairs).
• btumqa_225k/btumqa_225k_val.csv: The validation split used in the AUGR-VQA study.
• btumqa_225k/btumqa_225k_test.csv: The testing split used in the AUGR-VQA study.

Data Fields

Each CSV file contains the following 47 fields:

1. Core VQA Fields

• qa_id: Unique identifier for the question-answer pair (e.g., BTUMQA225K_24716f3aeba265).
• question: The natural language question regarding the MRI slice.
• answer: The target ground truth answer.
• split: Dataset split (e.g., train, val, test).
• difficulty_level: Question difficulty level (e.g., easy, medium, hard).
• answer_type: The type of answer (e.g., classification).

2. Image & Patient Metadata

• patient_id: Patient identifier from RSNA BraTS 2021.
• slice_id / image_id: MRI slice index number.
• unique_id: Patient and slice combined identifier (e.g., 00000_049).
• modality_context: Available MRI modalities (e.g., FLAIR|T1|T1ce|T2).

3. Question Metadata

• question_family: Semantic template family (e.g., confidence_qualified_presence, uncertainty_gap_bucket, dominant_region_under_uncertainty).
• question_style: Formatting style of the question (e.g., confidence_qualified, bucketed, ambiguity_sensitive).
• template_id / paraphrase_id: Identifiers for the underlying templates and natural language paraphrases.
• region_target_primary / region_target_secondary: Target brain regions referred to in the question (e.g., edema, enhancing, tumor).

4. Region Area Statistics (in Pixels)

• edema_area_pixels, ncr_net_area_pixels, enhancing_area_pixels, tumor_area_pixels, context_area_pixels, global_area_pixels: Surface area of each segmented region.

5. Uncertainty & Modulation Weight Estimates (MC-Dropout)

• edema_uncertainty, ncr_net_uncertainty, enhancing_uncertainty, tumor_uncertainty, context_uncertainty, global_uncertainty: MC-dropout pixel-level variance scores.
• edema_weight, ncr_net_weight, enhancing_weight, tumor_weight, context_weight, global_weight: Region weight coefficients used for token modulation.
• mc_dropout_passes: Number of MC-dropout forward passes used (e.g., 20).

6. Generated Intermediate Token File References

• phase2a_token_file, slice_index_in_token_file, phase2c_file, slice_index_in_ugtm_file: File paths and slice indices pointing to the extracted visual tokens and UGTM-modulated tokens. Note: The actual token files are not hosted in this repository. Instead, these paths represent where the files will be saved once dynamically generated by running the data processing pipelines in the official AUGR-VQA Code Repository (Private, publishing soon).

---

🚀 How to Use

You can load these CSV files directly from the subfolder using the Hugging Face datasets library:

from datasets import load_dataset

# 1. Load the training split
train_dataset = load_dataset(
    "shuvomonowar00/BTUMQA", 
    data_files="btumqa_225k/btumqa_225k_train.csv"
)

# 2. Load the validation split
val_dataset = load_dataset(
    "shuvomonowar00/BTUMQA", 
    data_files="btumqa_225k/btumqa_225k_val.csv"
)

# 3. Load the testing split
test_dataset = load_dataset(
    "shuvomonowar00/BTUMQA", 
    data_files="btumqa_225k/btumqa_225k_test.csv"
)

# 4. Load the full pool of QA pairs
full_dataset = load_dataset(
    "shuvomonowar00/BTUMQA", 
    data_files="btumqa_225k/btumqa_225k_qa_pairs.csv"
)