gokhankocmarli/inline-digital-holography-v4

Synthetic In-line Holograms of Cells v4 (1024 px, magnified)

Simulated in-line (Gabor) holograms of red blood cells and adherent cells, recorded through a 10x/0.45 microscope objective on a Sony IMX183 camera at 532 nm, with the complex object field as ground truth. The dataset holds a training set, a training set with recording-distance errors, an in-distribution test set, four out-of-distribution splits paired hologram by hologram with the test set, a dense split and a distance sweep: 24,000 holograms of 1024 x 1024 pixels in all.

Every set was generated with hologen from a recipe file in holopaswin recipes/, and each set folder contains the complete recipe.toml it was made from, so every hologram can be regenerated bit for bit from its seed.

• Curated by: Gökhan Koçmarlı (Gyokhan Kochmarla) | Google Scholar | GitHub
• License: MIT
• Generator: hologen, dataset format 2
• Size: 24,000 holograms, 99.7 GB

Sets

Folder	Split name	Holograms	Seeds	What differs from the test set
train	train	8,000	10000-17999	training distribution, nominal distance
train-z-augmented	train_z_augmented	8,000	10000-17999	same cells as train, recording distance drawn from 14-26 um
test	test	1,000	200000-200999	in-distribution test set
test-phase	test_phase	1,000	200000-200999	twice the phase of every cell
test-absorption	test_absorption	1,000	200000-200999	three times the haemoglobin absorption of training for red blood cells (rim transmission 0.94 -> 0.82); adherent cells unchanged
test-low-light	test_low_light	1,000	200000-200999	a quarter of the exposure, twice the relative shot noise
test-strong-speckle	test_strong_speckle	1,000	200000-200999	laser speckle contrast 0.25 instead of 0.12 (about twice)
test-dense	test_dense	400	300000-300399	twice the red blood cell density (501 cells, 40 % of the field) and an adherent coverage of 0.5 instead of 0.263 (other cells)
sweep/offset_DZ	sweep_offset_DZ	200 per offset	900000-900199	recording distance 20 + DZ um, DZ in 0, ±1, ±2, ±4, ±6, ±8, ±10 (m4 = -4, p4 = +4 in split names)

The phase, absorption, low-light and strong-speckle splits use the test set's seeds and groups, so hologram i of each holds the same cells as test hologram i and differs in one property. They keep the exposures calibrated for the training distribution on purpose, so some holograms saturate a few pixels more often than in training (fraction of holograms with any clipped pixel, largest clipped fraction of one hologram):

Set	Red blood cells	Adherent cells
train, test	0-4 %, at most 0.0003 %	2-8 %, at most 0.003 %
test_phase	100 %, at most 0.26 %	78-90 %, at most 0.024 %
test_strong_speckle (laser groups)	100 %, at most 0.0026 %	24 %, at most 0.0014 %
sweep_offset_p10	100 %, at most 0.05 %	0-12 %, at most 0.003 %

clipped_fraction in the metadata gives the value for every hologram.

Within each set the holograms are split equally over eight groups: four noise configurations times two object families. The noise levels below were measured with hologen measure on the training recipe; test_low_light has twice the relative shot noise and test_strong_speckle about twice the laser speckle.

Configuration	Noise	rms deviation from the noise-free hologram at the training exposures, red blood cells / adherent cells
ideal	none (12-bit quantisation only)	0.04 % / 0.05 %
camera	shot + dark current + read noise + 12-bit ADC	1.9 % / 2.3 %
camera_led_speckle	camera + speckle contrast 0.014	2.4 % / 2.7 %
camera_laser_speckle	camera + speckle contrast 0.12	12.9 % / 12.3 %

Simulation settings

Parameter	Value
Wavelength	532 nm
Objective	10x, NA 0.45 (coherent pupil cutoff applied on the way to the sensor)
Camera	Sony IMX183: 2.4 um pixels, 12 bit, saturation 14,837 e-, read noise 3.3 e- (EMVA 1288)
Pixel size in the object plane	0.24 um, field of view 245.8 x 245.8 um
Recording distance	20 um (nominal)
Background exposure	2,735 e- (red blood cells), 1,855 e- (adherent cells): per family, the largest background at which 95 % of noise-free laser-speckle holograms keep their brightest point plus three standard deviations of shot noise below saturation (400 calibration holograms per family)
Speckle grain size	0.59 um
Objects	red blood cells (0.00415 cells per um², 251 per field) or adherent cells (coverage 0.263, median 27 cells per field; placement stops after 200 attempts, so a few fields stay just below the target)

The full set of values, with their sources, is in each folder's recipe.toml.

Files

Each set folder holds, per batch of 250 holograms:

• hologram_XXXX.parquet: sample_id, intensity (1024 x 1024, uint16 camera counts).
• ground_truth_XXXX.parquet: sample_id, real and imag (1024 x 1024, float32), the complex object field relative to the illumination.
• metadata_XXXX.parquet: sample_id, config_name, object_family, seed, objects_placed, background_electrons, clipped_fraction, z_true and z_nominal (um).
• recipe.toml: the complete recipe the set was generated from.

Rows of the three files of one batch belong together by sample_id.

The true phase of the thickest adherent cells exceeds π (about 4.8 rad at the 95th percentile of peak optical path difference), and in test_phase both families exceed it, so angle(real + i imag) wraps. Compare complex fields, or unwrap, when computing phase metrics.

Loading

With HoloPASWIN, which reads the format and the recording distance:

from huggingface_hub import snapshot_download
from holopaswin.dataset import HoloDataset

root = snapshot_download("gokhankocmarli/inline-digital-holography-v4", repo_type="dataset", allow_patterns="test/*")
dataset = HoloDataset(f"{root}/test", target_size=224, img_dim=1024, return_z=True)
hologram, field, z = dataset[0]  # (1, 224, 224), (2, 224, 224) real and imaginary, z in m

With datasets, one config per file type and one split per set:

from datasets import load_dataset

metadata = load_dataset("gokhankocmarli/inline-digital-holography-v4", "metadata", split="test")
holograms = load_dataset("gokhankocmarli/inline-digital-holography-v4", "hologram", split="test", streaming=True)

Regenerating

hologen generate --recipe recipes/test.toml --output test

hologen generate warns for the phase, absorption, strong-speckle and dense recipes, because they change settings the exposures were calibrated for; this is intended.

Limitations

• The holograms are simulated. The cell models are procedural and calibrated to published optical path statistics, not traced from real cells.
• Adherent cells are pure phase objects in every set, as no measured absorption of stained adherent cells near 532 nm was available.
• Earlier versions (v1-v3, 224 px, ellipse objects, lensless geometry) are not compatible with this one.

Changes

• 2026-10-24: test-dense regenerated. The first upload requested three times the red blood cell density, which random placement without overlap cannot reach, and stopped placing cells early (about 1.5 times the training density and an adherent coverage of 0.37). It now holds twice the red blood cell density and an adherent coverage of 0.5, as stated above. Every other set is unchanged.
• 2026-10-22: first upload.

Citation

@misc{kochmarla2026synthetic_inline_holograms_of_cells_v4,
  author       = {Ko{\c{c}}marl{\i}, G{\"o}khan},
  title        = {Synthetic In-line Holograms of Cells v4 (1024 px)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/gokhankocmarli/inline-digital-holography-v4}}
}