USEnhance2023_Aligned

Pixel-aligned low-quality / high-quality ultrasound image pairs derived from the USEnhance2023 Grand Challenge. This dataset is a release artifact of the paper Style-Driven Data Synthesis and Degradation-Aware Enhancement for Ultrasound Image Restoration and is intended for academic research only.

Why this dataset

The raw USEnhance2023 corpus contains low_quality/ and high_quality/ ultrasound scans, but the two splits are acquired from different patients with different probes — they are unpaired. This blocks supervised pixel-level training objectives (L1 / L2 / LPIPS) that need ground-truth correspondence.

We use a CycleDiff (cycle-consistent latent diffusion) pipeline to translate each high-quality image into a pixel-aligned synthetic low-quality counterpart, producing the Aligned_LQ/ split. This makes USEnhance2023 directly usable for supervised image restoration / super-resolution / enhancement research.

Contents

Split	Count	Resolution	Bit-depth	Source
train/GT/	840	256×256	8-bit grayscale	Raw USEnhance2023 high-quality
train/LQ/	840	256×256	8-bit grayscale	Raw USEnhance2023 low-quality (real, unpaired)
train/Aligned_LQ/	840	256×256	8-bit grayscale	CycleDiff inference on train/GT/ (synthetic, pixel-aligned)
test/GT/	210	256×256	8-bit grayscale	Raw USEnhance2023 high-quality
test/LQ/	210	256×256	8-bit grayscale	Raw USEnhance2023 low-quality (real, unpaired)
test/Aligned_LQ/	210	256×256	8-bit grayscale	CycleDiff inference on test/GT/
test/LR_64/	210	64×64	8-bit grayscale	test/GT/ bicubic ↓4 (×4 SR benchmark)

File names are aligned across splits (e.g. train/GT/breast_1213.png ↔ train/Aligned_LQ/breast_1213.png).

Recommended use

• Supervised image restoration / enhancement — train with (Aligned_LQ, GT) pairs.
• Zero-shot / unpaired generalization eval — test on real-world LQ/ to measure sim-to-real gap.
• ×4 Super-resolution benchmark — use LR_64/ → GT/ (256×256).

Download

pip install -U "huggingface_hub[cli]"
hf download Jason0411202/USEnhance2023_Aligned \
    --repo-type dataset \
    --local-dir data_factory/USEnhance2023_Aligned

Or jointly with our model weights in a single command:

hf download Jason0411202/DDG_LoRA --local-dir .

(the model repo mirrors this dataset under data_factory/USEnhance2023_Aligned/).

Reproducing from raw

The full CycleDiff training + inference pipeline that produced this dataset is open-source at Jason0411202/ITRI-PiSA-SR. After obtaining raw USEnhance2023 from Grand Challenge:

./data_factory/build_USEnhance2023_Aligned.sh

License & attribution

This is a derivative dataset. The GT/ and LQ/ folders contain images from the original USEnhance2023 Grand Challenge dataset. The Aligned_LQ/ and LR_64/ folders are derivative artifacts generated by the authors.

• Use is restricted to academic / non-commercial research.
• Users must cite both the original USEnhance2023 challenge and this work.
• Redistribution beyond academic use requires permission from the original USEnhance2023 organizers (contact: guoyi@fudan.edu.cn).
• If the original organizers request removal, we will comply.

Citation

If you use this dataset, please cite:

@inproceedings{usenhance2023,
  title  = {Ultrasound Image Enhancement Challenge 2023},
  author = {USEnhance2023 Organizers},
  year   = {2023},
  url    = {https://ultrasoundenhance2023.grand-challenge.org/}
}

@misc{usenhance2023_aligned,
  title  = {USEnhance2023\_Aligned: Pixel-aligned LQ-HQ ultrasound pairs via CycleDiff},
  author = {Jason and collaborators},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Jason0411202/USEnhance2023_Aligned}}
}

Contact

Please open an issue on the Hugging Face dataset page or the GitHub repo.