---
pretty_name: Brain MRI Dataset with Tumor Masks and Coarse Captions
language:
- en
task_categories:
- image-segmentation
- image-to-text
tags:
- medical-imaging
- brain-mri
- tumor-segmentation
- segmentation-masks
- image-captioning
- radiology
size_categories:
- 10K<n<100K
configs:
- config_name: caption_pairs
  data_files:
  - split: train
    path: brain_mri_positive_caption_pairs.csv
---

# Brain MRI Dataset with Tumor Masks and Coarse Captions

## Dataset Summary

This dataset is a curated collection of 2D brain MRI images paired with tumor segmentation masks and positive coarse captions. It was assembled for medical vision-language and segmentation experiments, especially tasks that require linking an MRI image, its lesion mask, and a text description of the visible tumor morphology.

Each row in the main CSV file contains:

- an image path,
- a corresponding tumor mask path,
- a positive caption derived from the `coarse_description` field in the MM-NeuroOnco metadata.

The positive captions describe coarse tumor-related attributes such as tumor type, modality, approximate location, size, shape, and spread pattern when available.

## Main Annotation File

The main index file is:

```text
brain_mri_positive_caption_pairs.csv
```

It has the following columns:

| Column | Description |
|---|---|
| `image_path` | Relative path to the MRI image. Paths start with `Brain MRI DATASET/`. |
| `mask_path` | Relative path to the corresponding tumor segmentation mask. Paths start with `Brain MRI DATASET/`. |
| `caption` | Positive coarse caption for the image, extracted from MM-NeuroOnco `coarse_description`. |

The dataset contains 22,150 image-mask-caption rows. No missing caption mappings were found during CSV generation.

## Data Composition

| Source folder | Number of image-mask-caption pairs |
|---|---:|
| `BRISC2025_segmentation` | 4,793 |
| `BraTS16_MMNeuroOnco` | 5,000 |
| `Brain_Tumor_Classification_2D_masked` | 3,166 |
| `Figshare_from_mat_all` | 3,064 |
| `MMNeuroOnco_tumor_14_t1ce_mask` | 3,064 |
| `MMNeuroOnco_tumor_19_mask` | 3,063 |
| **Total** | **22,150** |

## Folder Structure

```text
Brain MRI DATASET/
  brain_mri_positive_caption_pairs.csv
  brain_mri_positive_caption_pairs_report.txt
  README.md

  BRISC2025_segmentation/
    train/
      images/
      masks/
    test/
      images/
      masks/

  BraTS16_MMNeuroOnco/
    16_t1_t2_t1ce_flair_mask/
      BraTS2021_*/
        *_flair.png
        *_t1.png
        *_t1ce.png
        *_t2.png
        *_seg.png

  Brain_Tumor_Classification_2D_masked/
    images/
      Astrocytoma/
      Glioblastoma/
      Gliomas/
      Meningioma/
      Neurocytoma/
    masks/
      Astrocytoma/
      Glioblastoma/
      Gliomas/
      Meningioma/
      Neurocytoma/

  Figshare_from_mat_all/
    images/
    masks/

  MMNeuroOnco_tumor_14_t1ce_mask/
    images/
    masks/

  MMNeuroOnco_tumor_19_mask/
    images/
    masks/
```

## Caption Construction

Captions were not newly generated from scratch for this dataset. Instead, they were mapped from the MM-NeuroOnco metadata file:

```text
MM-NeuroOnco-Images/metadata/Dataset_For_LLM_Step1.json
```

For each image, the corresponding metadata entry was located and its `coarse_description` field was used as the positive caption.

Example caption:

```text
A T1 brain MRI scan showing signs of Glioma. The mass is large/extensive and located in the center of the image. It presents as a lobulated structure and appears as a solitary lesion.
```

## Mapping Rules

The CSV was generated by matching local image files to their corresponding MM-NeuroOnco metadata entries. The following mapping rules were used:

| Local subset | Metadata mapping rule |
|---|---|
| `BRISC2025_segmentation` | Matched to `tumor/05_t1_mask/<same_filename>` |
| `BraTS16_MMNeuroOnco` | Matched to `glioma/16_t1_t2_t1ce_flair_mask/<case>/<modality_image>` |
| `MMNeuroOnco_tumor_14_t1ce_mask` | Matched to `tumor/14_t1ce_mask/<index>.png` |
| `Figshare_from_mat_all` | Matched to the same numbered entries as `tumor/14_t1ce_mask/<index>.png` |
| `MMNeuroOnco_tumor_19_mask` | Matched to `tumor/19_mask/images/<same_filename>` |
| `Brain_Tumor_Classification_2D_masked` | Matched to source-10 MM-NeuroOnco metadata by tumor class and filename |

A generation report is included here:

```text
brain_mri_positive_caption_pairs_report.txt
```

It records the final row count and confirms that the number of missing caption mappings is zero.

## Intended Uses

This dataset is intended for research experiments such as:

- tumor segmentation from brain MRI images,
- vision-language alignment using image-mask-caption triples,
- DPO or preference-learning experiments with medical image-text pairs,
- weakly supervised or prompt-guided segmentation,
- evaluation of whether visual encoders preserve lesion-relevant information.

## Example Usage

```python
import pandas as pd
from pathlib import Path

repo_root = Path('/path/to/project')
csv_path = repo_root / 'Brain MRI DATASET' / 'brain_mri_positive_caption_pairs.csv'

df = pd.read_csv(csv_path)

row = df.iloc[0]
image_path = repo_root / row['image_path']
mask_path = repo_root / row['mask_path']
caption = row['caption']

print(image_path)
print(mask_path)
print(caption)
```

## Notes and Limitations

- The captions are coarse descriptions and should not be interpreted as full radiology reports.
- The captions are positive tumor descriptions derived from metadata, not free-form clinical diagnoses.
- Some subsets may share the same underlying image content or metadata source, especially the Figshare-derived and MM-NeuroOnco `14_t1ce_mask` subsets.
- The masks come from the available source annotations or converted source files; mask conventions may differ across subsets.
- This dataset is intended for research use only and must not be used for clinical decision-making.

## Ethical and Safety Considerations

This dataset contains medical imaging data. Models trained on it may produce incorrect, incomplete, or misleading outputs. Any model or analysis built using this dataset should be validated carefully and should not be deployed in clinical settings without appropriate regulatory review, expert validation, and institutional approval.

## Provenance

This collection aggregates image-mask pairs prepared from several public or locally reconstructed brain MRI segmentation sources and aligns them with coarse captions from MM-NeuroOnco metadata. Users should review the licensing and usage terms of the original data sources before redistributing or using the dataset in downstream projects.