---
license: cc-by-nc-4.0
task_categories:
- image-to-image
tags:
- medical
- CT
- flow-matching
- image-translation
- phase-translation
- nifti
---

# MS-Thesis Dataset (CT Phase Translation Dataset for Flow Matching)

## Dataset Description

This dataset contains paired CT scans for training Flow Matching models on CT phase translation tasks. 
The dataset was created as part of a Master's degree thesis focused on translating between native (non-contrast) and arterial phase CT images using generative flow-based models.

### Dataset Summary

- **Task**: CT phase translation (native -> arterial phase)
- **Modality**: Computed Tomography (CT)
- **File Format**: NIfTI (.nii.gz)
- **Registration**: ANTs (Advanced Normalization Tools) with `antsRegistrationSyN[so]` transform
- **Spatial Alignment**: All images are registered in original spacing.
- **Time**: It took about 2 hours of time on a 128-core CPU server to register each of the images.


## Dataset Structure
- **art.nii.gz**: Arterial phase CT scan
- **nat.nii.gz**: Native phase CT scan  
- **body_mask.nii.gz**: Body segmentation mask

### Splits

| Split      | Number of Cases | Description |
|------------|-----------------|-------------|
| Train      | 80 cases        | Training set |
| Test       | 20 cases        | Testing set |
| Validation | 20 cases        | Validation set |

## Data Collection and Preprocessing

### Registration Pipeline

All images underwent spatial registration using ANTs (Advanced Normalization Tools):

1. **Fixed Image**: Native phase CT (`nat.nii.gz`)
2. **Moving Image**: Arterial phase CT (before registration)
3. **Transform Type**: `antsRegistrationSyN[so]` (symmetric normalization with small deformation)
4. **Fixed Image Mask**: Body mask applied during registration to focus alignment on anatomical structures

### Preprocessing Steps

1. **Body Mask Creation**: Automatic segmentation of body region from background
2. **Dust Removal**: Both native and arterial phases were cleaned using the body mask to eliminate:
   - External objects (clothing, table artifacts)
   - Scanner noise outside body region
   - Non-anatomical intensities
3. **Purpose**: Reduces domain shift and focuses model learning on relevant anatomical transformations.

### Quality Control

- All registrations were verified for anatomical alignment.
- Images are free of major artifacts that would render them undiagnostic.
- Intensity values preserved within body mask region.
- The values on the border of the images that appeared after registration have been removed.