OR-PAM-Reg-4K: A Benchmark Dataset for Bidirectional OR-PAM Registration

Overview

OR-PAM-Reg-4K is a benchmark dataset for evaluating image registration methods in bidirectional optical-resolution photoacoustic microscopy (OR-PAM). The dataset contains 4,248 paired image samples (512 × 512 pixels) of in vivo mouse brain vasculature, with separated odd-line (forward scan) and even-line (backward scan) acquisitions.

Dataset Statistics

Split	Samples	Percentage
Train	3,396	80%
Val	420	10%
Test	432	10%
Total	4,248	100%

Data Format

OR-PAM-Reg-4K/
├── train.h5    # Training set (1.5 GB)
├── val.h5      # Validation set (193 MB)
├── test.h5     # Test set (200 MB)
└── README.md   # This file

HDF5 Structure

Each .h5 file contains:

Key	Shape	Dtype	Description
odd	(N, 1, 512, 256)	float32	Odd-line (forward scan) images
even	(N, 1, 512, 256)	float32	Even-line (backward scan) images
sample_ids	(N,)	string	Unique sample identifiers

• Pixel values: Normalized to [0, 1]
• Channel: Single-channel grayscale (photoacoustic signal intensity)
• Dimensions: (batch, channel, height, width) in PyTorch convention

Quick Start

Loading Data (Python)

import h5py
import torch
from torch.utils.data import Dataset

class ORPAMRegDataset(Dataset):
    def __init__(self, h5_path):
        with h5py.File(h5_path, 'r') as f:
            self.odd = f['odd'][:]
            self.even = f['even'][:]

    def __len__(self):
        return len(self.odd)

    def __getitem__(self, idx):
        odd = torch.from_numpy(self.odd[idx]).float()
        even = torch.from_numpy(self.even[idx]).float()
        return odd, even

# Usage
train_dataset = ORPAMRegDataset('train.h5')
odd, even = train_dataset[0]
print(f'Odd shape: {odd.shape}, Even shape: {even.shape}')
# Output: Odd shape: torch.Size([1, 512, 256]), Even shape: torch.Size([1, 512, 256])

Memory-Efficient Loading

For large-scale training, use memory-mapped loading:

import numpy as np

class ORPAMRegDatasetMmap(Dataset):
    def __init__(self, data_dir, split='train'):
        # First convert h5 to npy if needed
        self.odd = np.load(f'{data_dir}/{split}_odd.npy', mmap_mode='r')
        self.even = np.load(f'{data_dir}/{split}_even.npy', mmap_mode='r')

    def __getitem__(self, idx):
        return (torch.from_numpy(self.odd[idx].copy()).float(),
                torch.from_numpy(self.even[idx].copy()).float())

Benchmark Task

Goal: Register even-line images to odd-line images, correcting both domain shift and geometric misalignment.

Evaluation Metrics

Metric	Description	Range
NCC	Normalized Cross-Correlation	[-1, 1], higher is better
SSIM	Structural Similarity Index	[0, 1], higher is better
PSNR	Peak Signal-to-Noise Ratio	dB, higher is better

Baseline Results (Author Implementation)

The following baselines are implemented and evaluated by the dataset authors on the test set (432 samples).

Traditional Methods

Method	SSIM	PSNR	NCC
Unregistered	0.482	19.46	0.167
SIFT	0.679	24.14	0.723
Demons	0.579	20.35	0.323
Optical Flow	0.455	18.99	0.061
SyN (ANTs)	0.613	21.55	0.411

Deep Learning Methods

Method	SSIM	PSNR	NCC
VoxelMorph	0.659	22.88	0.724
TransMorph	0.641	21.09	0.594

Published Methods (Third-Party Comparison)

Results reported in the original publications and our reproduction on OR-PAM-Reg-4K. We welcome community contributions to this leaderboard.

Method	Paper	Reported SSIM	Reported NCC	Reproduced SSIM	Reproduced NCC
—	—	To be updated	To be updated	To be updated	To be updated

If you have evaluated your method on OR-PAM-Reg-4K, please open an issue or contact us to add your results.

Citation

If you use this dataset, please cite:

@misc{tianyan_zhang_2026,
  author    = {Tianyan Zhang and Chengliu Yan and Xiangzhi Lan},
  title     = {OR-PAM-Reg-4K: A Benchmark Dataset for Bidirectional OR-PAM Registration},
  year      = {2026},
  url       = {https://huggingface.co/datasets/chengliuyan/OR-PAM-Reg-4K},
  doi       = {10.57967/hf/7721},
  publisher = {Hugging Face}
}

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).