MOSXAV: A Benchmark Dataset for Multi-Object Segmentation in X-ray Angiography Videos

1. Overview

MOSXAV is a benchmark dataset designed for multi-object segmentation in X-ray angiography videos. It provides high-quality, manually annotated segmentation ground truth, supporting the analysis of vascular structures in dynamic medical imaging. Each video contains 33~70 frames at a resolution of 512×512 pixels. Vascular regions are annotated by experienced radiologists, with annotations focused on one or two key frames where the contrast agent is most prominent.

• The training and validation sets include 30 sequences (2,335 frames), with annotations every 5 frames.
• The test set consists of 12 sequences (488 frames), with frame-level annotations throughout.

MOSXAV provides a valuable resource for the development and benchmarking of methods in X-ray angiography video segmentation.

2. Annotation Protocol

To ensure high-quality and anatomically accurate labels, we implemented a rigorous, multi-stage annotation workflow. This process combined the efficiency of deep learning with the precision of manual expert refinement. The protocol consisted of four primary phases:

• Annotator Training: Annotators were trained on a specialized subset of images to standardize their understanding of anatomical structures and specific labeling guidelines.
• Semi-Automated Initialization: We utilized a semi-automated approach to generate initial segmentation masks. This was powered by the PaddleSeg framework, leveraging models pre-trained on extensive image and video datasets to provide a robust baseline.
• Expert Revision: Human annotators meticulously reviewed the AI-generated masks. This involved careful delineation of vessel boundaries and manual adjustments to correct any discrepancies in the automated output.
• Consensus & Quality Assurance: To maintain consistency, a final review and consensus-building phase were conducted, ensuring that all labels met our strict quality benchmarks.

3 Object Categories and Statistics

The MOSXAV dataset is designed to support two distinct medical imaging challenges: Video Object Segmentation (VOS) and Multi-class Semantic Segmentation.

3.1 Video Object Segmentation (VOS)

The VOS task focuses on the temporal tracking and segmentation of coronary arteries as they are opacified by contrast agents. This task requires high temporal consistency across video sequences:

• Objective: Segmenting coronary arteries filled with contrast agents throughout the cardiac cycle.
• Object Density: Train & Val Sets: Up to 5 individual objects per sequence. Test Set: Increased complexity with up to 10 individual objects per sequence to evaluate model scalability and robustness.

3.2 Semantic Segmentation

The semantic segmentation task targets the simultaneous identification of critical intervention tools and anatomical features. We define four primary categories:

3.3 File Structure

The MOSXAV dataset is organized into a hierarchical directory structure to support both video-level (VOS) and frame-level (Semantic Segmentation) tasks.

MOSXAV_Dataset/
├── trainval/
│   ├── Annotations/                # VOS instance masks (unique ID per branch)
│   │   └── v00/                    # Sequence folder
│   │       ├── 00000.png           # Frame-wise instance mask
│   │       └── ...
│   ├── Annotations_Semantic/       # Multi-class semantic masks (Label IDs 0-4)
│   │   └── v00/
│   │       ├── 00000.png
│   │       └── ...
│   ├── JPEGImages/                 # Raw X-ray Angiography frames
│   │   └── v00/
│   │       ├── 00000.jpg
│   │       └── ...
│   ├── ImageSets/                  # Split lists and first-frame metadata
│   │   ├── train.txt
│   │   ├── val.txt
│   │   └── val_first_mask.json     # Frame ID of each object's first appearance
│   └── labels.json                 # Global category metadata
└── test/
    ├── Annotations/
    ├── Annotations_Semantic/       # Multi-class semantic masks (Label IDs 0-5)
    ├── JPEGImages/
    ├── ImageSets/
    │   ├── test.txt
    │   └── test_first_mask.json    # Frame ID of each object's first appearance, along with the seen and unseen object classes in the training set
    └── labels.json

4. Download

The MOSXAV Dataset is hosted across multiple cloud storage platforms to ensure accessibility and high download speeds globally.

Source	Download Link	Extraction Code / Notes
OneDrive	Link
Google Drive	Link	-
Baidu Pan	Link
Hugging Face	Link	-

5. License

The dataset is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. See LICENSE for details.

Contributors

• Principal investigator: YingLiangUEA
• Core contributors: xilin-x, YingLiangUEA
• Student contributors: EthanKoland

Citation

Please consider citing MOSXAV if it helps your research.

@article{FSVOSXA,
            title={Few-Shot Video Object Segmentation in X-Ray Angiography Using Local Matching and Spatio-Temporal Consistency Loss},
            author={Xi, Lin and Ma, Yingliang and Zhuang, Xiahai},
            journal={arXiv preprint arXiv:2601.00988},
            year={2026}
        }

@InProceedings{RNPLL,
            author={Xi, Lin and Ma, Yingliang and Wang, Cheng and Howell, Sandra and Rinaldi, Aldo and Rhode, Kawal S.},
            title={Robust Noisy Pseudo-Label Learning for Semi-supervised Medical Image Segmentation Using Diffusion Model},
            booktitle={Deep Generative Models Workshop, International Conference on Medical Image Computing and Computer-Assisted Intervention (DGM4MICCAI)},
            year={2026},
            pages={12--23}
        }

Contact

For questions or feedback, please contact:

• Lin Xi:
• Yingliang Ma: