WafaaFraih/rocov2-modality-splits

ROCOv2 Modality-Specific Dataset Splits

Dataset Description

This dataset contains modality-specific splits of the ROCOv2 radiology dataset, organized and processed for training specialized medical image captioning models.

Dataset Summary

• Total Samples: 1,000
• Modalities: 5
• Splits per Modality: train, validation, test
• Random Seed: 42
• Processing Date: 2025-08-31 12:52:59.233482

Modality Distribution

Modality	Samples	Percentage
CT	188	18.8%
X-ray	199	19.9%
MRI	200	20.0%
Microscopy	207	20.7%
Ultrasound	206	20.6%

Dataset Structure

Data Fields

• caption (string): Medical description of the image
• modality (string): Imaging modality (CT, MRI, X-ray, Ultrasound, Microscopy, Other)
• modality_id (int): Numerical ID for the modality
• caption_length (int): Number of words in the caption
• length_category (string): Caption length category (short, medium, long)
• original_index (int): Index in the original dataset
• image (Image): Medical image (if available)

Data Splits

Each modality is split into:

• Train: ~80% for model training
• Validation: ~10% for model validation
• Test: ~10% for final evaluation

Split Statistics

CT

• train: 150 samples
• val: 18 samples
• test: 20 samples

X-ray

• train: 159 samples
• val: 19 samples
• test: 21 samples

MRI

• train: 160 samples
• val: 20 samples
• test: 20 samples

Microscopy

• train: 165 samples
• val: 20 samples
• test: 22 samples

Ultrasound

• train: 164 samples
• val: 20 samples
• test: 22 samples

Usage

from datasets import load_dataset

# Load a specific modality
ct_dataset = load_dataset("WafaaFraih/rocov2-modality-splits", "CT")

# Load all modalities
all_datasets = {}
modalities = ["CT", "MRI", "X-ray", "Ultrasound", "Microscopy", "Other"]

for modality in modalities:
    try:
        all_datasets[modality] = load_dataset("WafaaFraih/rocov2-modality-splits", modality)
        print(f"Loaded {modality}: {len(all_datasets[modality]['train'])} train samples")
    except:
        print(f"{modality} not available")

# Access specific split
train_data = ct_dataset['train']
val_data = ct_dataset['validation'] 
test_data = ct_dataset['test']

Modality-Specific Characteristics

CT (Computed Tomography)

• Keywords: axial, coronal, sagittal, contrast, enhancement, density, hounsfield
• Typical caption prefixes: "This CT scan shows", "CT imaging reveals"

MRI (Magnetic Resonance Imaging)

• Keywords: t1, t2, flair, dwi, gadolinium, signal, intensity, weighted
• Typical caption prefixes: "This MRI shows", "Magnetic resonance imaging reveals"

X-ray (Radiography)

• Keywords: ap, lateral, pa, portable, upright, supine, opacity, lucency
• Typical caption prefixes: "This chest X-ray shows", "Radiographic examination reveals"

Ultrasound (Sonography)

• Keywords: echo, doppler, transducer, frequency, acoustic, anechoic, hyperechoic
• Typical caption prefixes: "This ultrasound shows", "Sonographic examination reveals"

Microscopy (Histopathology)

• Keywords: cellular, nuclear, cytoplasm, tissue, staining, morphology, biopsy
• Typical caption prefixes: "This microscopic image shows", "Histological examination reveals"

Training Recommendations

Modality-Specific Training

• Each modality has optimized hyperparameters
• Use modality-specific data augmentation
• Consider modality-specific evaluation metrics

Batch Sizes (recommended)

• CT: 8
• MRI: 6
• X-ray: 10
• Ultrasound: 8
• Microscopy: 6

Learning Rates (recommended)

• CT: 2e-5
• MRI: 1.5e-5
• X-ray: 3e-5
• Ultrasound: 2.5e-5
• Microscopy: 1.8e-5

Citation

If you use this dataset, please cite the original ROCOv2 paper and this processed version:

@article{rocov2,
  title={ROCOv2: Radiology Objects in COntext Version 2},
  author={...},
  journal={...},
  year={2023}
}

License

Please refer to the original ROCOv2 dataset license terms.

Dataset Creation

This dataset was created by processing the original ROCOv2 dataset with stratified splitting to ensure balanced representation across different caption lengths and modalities.

Processing Steps:

1. Modality classification based on caption content
2. Stratified splitting by caption length categories
3. Quality validation and metadata generation
4. Conversion to Hugging Face Dataset format

Contact

For questions about this processed dataset, please open an issue in the repository.