PASD — Placenta Accreta Spectrum MRI Dataset

A 3D MRI dataset for Placenta Accreta Spectrum (PAS) diagnosis with voxel-level lesion masks and case-level diagnostic labels. This dataset accompanies the paper:

3D Segment Anything Model with Visual Mamba for Diagnosing Placenta Accreta Spectrum, IEEE Transactions on Image Processing.

Source code for the proposed 3DSAMba method: https://github.com/Drchip61/PASD.

Dataset Summary

Split	Cases	Negative (label=0)	Positive (label=1)
train	184	61	123
test	60	20	40
total	244	81	163

Each case contains a single transverse-plane T2-weighted MRI volume of the uterus and the corresponding binary segmentation mask covering the suspected lesion region. Volumes are saved as NIfTI files (.nii.gz) at their native resolution; typical shape is (560, 560, ~55-70) with float64 intensities in roughly [0, 3500].

Files & Layout

PASD/
├── train/
│   ├── PASD_00001_1/
│   │   ├── PASD_00001_1_image.nii.gz   # MRI volume
│   │   └── mask.nii.gz                 # binary segmentation mask
│   ├── PASD_00002_1/
│   │   └── ...
│   └── PASD_00184_1/
└── test/
    ├── PASD_00185_1/
    │   └── ...
    └── PASD_00244_0/

• The directory name encodes the case id and the case-level class label (PASD_<5-digit-id>_<label>), where label ∈ {0, 1} indicates PAS-negative or PAS-positive respectively.
• Inside every case directory there is exactly one MRI volume (*_image.nii.gz) and one segmentation mask (mask.nii.gz).

This layout is the one expected by the dataloaders in the reference implementation. The classifier-stage dataset_class.py additionally reads predicted masks from a sibling directory (test_other/) — see the repository for details.

Privacy / De-identification

All cases have been fully de-identified:

• Original patient-name pinyin and hospital sequence numbers have been removed from both directory names and file names.
• NIfTI header fields that could contain free text (descrip, intent_name, aux_file, db_name) are emptied. They were already empty in the source data, but we scrub them defensively.
• No DICOM tags, accession numbers, or acquisition timestamps are distributed with the dataset.

The internal mapping between original case identifiers and the released PASD_xxxxx ids is not part of this release and is kept only by the data custodians.

How to Load

Plain PyTorch

import os
import nibabel as nib

CASE_DIR = "PASD/train/PASD_00001_1"

mri = nib.load(os.path.join(CASE_DIR, "PASD_00001_1_image.nii.gz")).get_fdata()
msk = nib.load(os.path.join(CASE_DIR, "mask.nii.gz")).get_fdata()

label = int(CASE_DIR[-1])    # 0 = PAS-negative, 1 = PAS-positive
print(mri.shape, msk.shape, label)

Hugging Face Datasets

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="ChipYTY/PASD",
    repo_type="dataset",
)

After the snapshot is available locally, the train/ and test/ folders can be plugged directly into the reference implementation's dataset.py.

Intended Use

• Lesion segmentation on placenta-region MRI.
• PAS positive vs. negative classification.
• Multi-task learning that couples segmentation and diagnosis.

The dataset is intended for research purposes only. It is not a substitute for clinical judgement and should not be used to make individual diagnoses.

Citation

@article{zhang2025pasd,
  title   = {3D Segment Anything Model with Visual Mamba for Diagnosing Placenta Accreta Spectrum},
  author  = {Zhang, Yuliang and He, Fang and Peng, Lulu and Guo, Qing and Yu, Lin and
             Wang, Zhijian and Shun, Wei and Liu, Jue and Chen, Yonglu and Huang, Jianwei and
             Bao, Zeye and Cai, Zhishan and Chen, Yanhong and Hu, Miao and Gu, Zhongjia and
             Shi, Yiyu and Yan, Tianyu and Zhang, Pingping and Ting, Song and Du, Lili and Chen, Dunjin},
  journal = {IEEE Transactions on Image Processing},
  year    = {2025}
}

License

Released under the MIT License.