KMH158/brats2020_shagufta

BraTS 2020 - Brain Tumor Radiology Report Generation Dataset

Dataset Description

This dataset contains paired brain MRI scans and radiology reports for training image-to-text models for automated radiology report generation.

Dataset Summary

• Total Patients: 166
• Total Images: 2080
• Modalities: T1, T1ce, T2, FLAIR (4 modalities per frame)
• Frames per Patient: 10 (optimal slice ± neighbors)
• Image Size: 224x224
• Source: BraTS 2020 Challenge

Splits

Split	Patients	Images
Train	166	1660
Validation	21	210
Test	21	210

Dataset Structure

Data Fields

• patient_id (int): Unique patient identifier
• report_number (str): Original report number (TR001, TR002, etc.)
• report (str): Full radiology report text
• slice_index (int): Axial slice index in the MRI volume
• frame_index (int): Frame number (0-9, where frame ~5 is optimal)
• is_optimal (bool): Whether this is the optimal slice for this patient
• optimal_slice (int): The optimal slice index for this patient
• image_t1 (Image): T1-weighted MRI image
• image_t1ce (Image): T1 contrast-enhanced MRI image
• image_t2 (Image): T2-weighted MRI image
• image_flair (Image): FLAIR MRI image

Data Sample

from datasets import load_dataset

dataset = load_dataset("KMH158/brats2020_shagufta")

# Access a sample
sample = dataset['train'][0]

print("Patient ID:", sample['patient_id'])
print("Report:", sample['report'][:200])
print("Modalities available:", ['t1', 't1ce', 't2', 'flair'])

# Display images
from PIL import Image
import matplotlib.pyplot as plt

fig, axes = plt.subplots(1, 4, figsize=(16, 4))
for idx, modality in enumerate(['t1', 't1ce', 't2', 'flair']):
    axes[idx].imshow(sample[f'image_{modality}'], cmap='gray')
    axes[idx].set_title(modality.upper())
    axes[idx].axis('off')
plt.show()

Methodology

Slice Selection

The optimal slice for each patient was determined using a trained regression model that:

1. Predicts tumor burden based on image features
2. Uses weighted scoring: 3×Necrotic + 2×Edema + 1×Enhancing
3. Selects the slice with maximum tumor burden

Then, 5 slices before and 4 slices after the optimal slice are extracted for data augmentation.

Image Preprocessing

1. Normalization: Global min-max normalization to [0, 255]
2. CLAHE: Contrast Limited Adaptive Histogram Equalization
3. Resizing: Bicubic interpolation to 224x224
4. Format: RGB (3-channel grayscale)

Report Mapping

Reports were mapped from the original CSV using the "Number" column (TR001, TR002, etc.) to BraTS patient IDs.

Intended Use

Primary Use Case

Training vision-language models for:

• Automated radiology report generation
• Medical image captioning
• Clinical decision support

Example Models

• BLIP (Salesforce/blip-image-captioning-base)
• GIT (microsoft/git-base)
• ViT + GPT-2
• Custom vision-language architectures

Example Training Code

from datasets import load_dataset
from transformers import BlipProcessor, BlipForConditionalGeneration
from torch.utils.data import DataLoader

# Load dataset
dataset = load_dataset("KMH158/brats2020_shagufta")

# Initialize model
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")

# Create dataloader
def collate_fn(batch):
    # Use T1ce modality (or any other)
    images = [item['image_t1ce'] for item in batch]
    texts = [item['report'] for item in batch]
    
    return processor(images=images, text=texts, return_tensors="pt", padding=True)

train_loader = DataLoader(dataset['train'], batch_size=8, collate_fn=collate_fn)

# Training loop
for batch in train_loader:
    outputs = model(**batch, labels=batch['input_ids'])
    loss = outputs.loss
    # ... training code

Multi-Modal Usage

This dataset includes all 4 MRI modalities. You can:

1. Single modality: Use T1ce (shows enhancement best)
2. Multi-modal: Concatenate or ensemble all 4 modalities
3. Modality-specific: Train separate models per modality

# Example: Multi-modal input
sample = dataset['train'][0]

# Stack all modalities
import numpy as np
multi_modal = np.stack([
    np.array(sample['image_t1']),
    np.array(sample['image_t1ce']),
    np.array(sample['image_t2']),
    np.array(sample['image_flair'])
], axis=-1)  # Shape: (224, 224, 12) - 4 modalities × 3 RGB channels

Limitations

• Dataset size: 166 patients (relatively small for deep learning)
• Single institution/protocol
• Expert annotations needed for validation
• Bias towards GBM cases in BraTS dataset
• English reports only

Citation

If you use this dataset, please cite:

@dataset{brats_radiology_reports_2024,
  title={BraTS 2020 Radiology Report Generation Dataset},
  author={Your Name},
  year={2024},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/KMH158/brats2020_shagufta}}
}

@article{menze2015multimodal,
  title={The multimodal brain tumor image segmentation benchmark (BRATS)},
  author={Menze, Bjoern H and Jakab, Andras and Bauer, Stefan and others},
  journal={IEEE transactions on medical imaging},
  volume={34},
  number={10},
  pages={1993--2024},
  year={2015},
  publisher={IEEE}
}

License

CC-BY-NC 4.0 (Non-commercial use only)

Acknowledgments

• BraTS Challenge organizers
• Original report annotators
• HuggingFace for dataset hosting