Hugging Face's logo

Datasets:
Flying-fish
/
MUST-ABUS

Tasks:
Image-to-Image
Languages:
English
Chinese
Size:
10K<n<100K
Tags:
medical-imaging
ultrasound
breast-ultrasound
image-registration
image-stitching
benchmark
DOI:
License:
Dataset card Files
xet
 Community
1
MUST-ABUS / README.md
Flying-fish's picture
Flying-fish
Publish reusable MUST-ABUS evaluation metrics
eacfd4d verified
preview code
|
Raw History Blame Contribute Delete
7.1 kB
metadata 
license: cc-by-nc-sa-4.0
task_categories:
  - image-to-image
language:
  - en
  - zh
tags:
  - medical-imaging
  - ultrasound
  - breast-ultrasound
  - image-registration
  - image-stitching
  - benchmark
pretty_name: >-
  MUST-ABUS: Multi-View Automated Breast Ultrasound Registration and Stitching
  Benchmark
size_categories:
  - 10K<n<100K

MUST-ABUS: Multi-View Automated Breast Ultrasound Registration and Stitching Benchmark

Dataset Summary

MUST-ABUS is a registration-centered benchmark for multi-view Automated Breast Ultrasound (ABUS) under partial-overlap and slice-level misalignment conditions. It supports unsupervised AP--MED and AP--LAT cross-view image registration, with registration-guided stitching used as downstream validation.

The cohort contains 356 retained view-specific 3D ABUS volumes from 90 subjects examined between 2020 and 2022. From these volumes, 25,692 cross-view 2D image pairs were constructed. The release includes de-identified retained source volumes, preprocessed image pairs, valid-region masks, pair correspondence files, and official patient-level splits.

Supported Tasks

  • Cross-view ABUS image registration
  • Cross-view correspondence estimation
  • Registration-guided image stitching
  • Partial-overlap-aware registration evaluation

MUST-ABUS is not designed for lesion detection, tumor segmentation, or diagnostic classification.

Dataset Structure

Source volumes

  • 90 subjects
  • 356 retained 3D view volumes: 136 AP, 126 LAT, and 94 MED
  • 330 frames per volume, with each frame measuring 482 x 841 pixels
  • GE Healthcare Invenia ABUS system with a C15-6XW transducer

Derived benchmark pairs

  • 25,692 AP--MED and AP--LAT image pairs
  • 10,744 AP--MED pairs
  • 14,948 AP--LAT pairs
  • Valid-region masks and pair-level metadata

Official patient-level split

Split Subjects Image pairs Retained volumes
Train 72 20,554 297
Validation 9 2,569 31
Test 9 2,569 28
Total 90 25,692 356

All data from the same subject are assigned to one split to prevent subject-level leakage.

Annotations and Metadata

Each benchmark sample contains a partially overlapping AP--MED or AP--LAT slice pair, corresponding valid-region masks, and pair metadata. Released metadata include anonymized subject identifiers, breast side, view labels, slice identifiers, volume dimensions, and non-identifying acquisition parameters.

The dataset does not include manually annotated dense deformation fields, pixel-level stitched ground truth, or lesion-level diagnostic annotations. Valid-region masks identify effective ultrasound imaging regions and are not tumor or lesion masks.

Manual landmark annotations (2026-10-16)

Download landmarks/landmarks.zip and extract it into the dataset root. The current release contains 8 processed CSV files, with 12,919 primary landmark correspondences across 1,164 image pairs from 6 test subjects and 7 primary volume-pair sets. An independent second-expert annotation of one shared case contributes 1,550 additional correspondences across 155 image pairs. Both experts are grouped under the same case:

landmarks/test/c79296d0-6347-4b92-9191-350bdcf443c7/right/ap-med/
  expert-1/processed_manual_landmark_pairs.csv
  expert-2/processed_manual_landmark_pairs.csv

Expert attribution follows curator confirmation: expert 1 annotated the primary sets for 01137f00-1912-466c-a5f4-3133297cf780, 34a9bfd9-4992-473e-88b9-6d31d1fc4263, 73dd6a2c-f6f4-49f1-bbc9-e4565aeb4fc5, and c79296d0-6347-4b92-9191-350bdcf443c7 (7,453 correspondences). Expert 2 annotated e5fbd376-8f52-4838-be34-451a97be6d26, f58efa8b-c199-4744-8f48-20950b8c1765, and the independent second set for the shared case (7,016 correspondences).

The curator inspected the landmarks in 3D Slicer and confirmed anatomical correspondence. All processed rows are retained. The archive includes no raw Slicer JSONs, and the four source-path columns ap_nifti_path, other_nifti_path, ap_annotation_file, and other_annotation_file have been removed. Other CSV columns and coordinates are preserved. The LPS x/y coordinates are in millimeters; *_position_z_index is a slice-table ordinal, not physical depth.

See the landmark guide, coverage and file manifest, and archive SHA-256 for details. Use the current ZIP and manifest together; this revision replaces the earlier filtered package. Existing dataset licensing and access terms apply.

Evaluation Code

The evaluation source for MSE, PSNR, SSIM, MI, Ghosting, and physical landmark TRE is available in evaluation/mustabus_metrics. See the English instructions or 中文说明.

The package includes geometric valid-region handling, invalid-output and transformation-failure handling, physical-coordinate conversion and pullback inversion for TRE, common-valid landmark selection, and patient-level aggregation. It runs independently of registration networks and checkpoints.

After downloading the source, open a terminal in evaluation/mustabus_metrics/:

uv sync --extra nifti
uv run python -m mustabus_metrics.examples.quickstart
uv run python -m unittest discover -s tests -v

Use the processed landmark release for TRE inputs. Registration outputs must supply the final fixed-to-moving sampling map and consistent physical image geometry. The example uses synthetic inputs; it does not run a registration model or automatically reproduce manuscript tables. Tests that require the original research files are skipped in this standalone source release.

Ethics, Privacy, and Intended Use

This retrospective single-center study was approved by the appropriate institutional Ethics Review Committee, which waived written informed consent. Direct identifiers are removed or replaced, anonymous study identifiers are used, and image content is checked for burned-in identifying information before release.

MUST-ABUS is intended for academic research on image registration, cross-view correspondence estimation, registration-guided stitching, and related medical-image analysis. Users must comply with the dataset license and data-use terms and must not attempt to re-identify any subject.

Limitations

  • The dataset is from a single clinical institution and one acquisition system.
  • Breast-side and view availability vary across subjects.
  • Adjacent slices from the same volume are correlated and should not be treated as independent clinical cases.
  • The dataset provides no dense deformation ground truth or lesion-level diagnostic labels.

Citation

Citation information and the persistent dataset DOI will be added when the benchmark paper and versioned archive record are finalized.