---
pretty_name: scientific_data_2026_masks
tags:
  - microscopy
  - mother-machine
  - bacteria
  - segmentation
  - instance-segmentation
  - zarr
---

#  Mask data for Hardo, Li, and Bakshi, 2026

## Dataset summary

This repository contains the released multi-hypothesis instance segmentation masks for the paper "An annotated timelapse imaging dataset on dormancy exit dynamics of *Escherichia coli* cells in Mother Machine".

The release consists of one Zarr store:

- `20260307_SB7_exit_snake_V4_1.segmentation_masks_multi_epoch_uint8_masks_only.zarr`

The store occupies approximately `4.6 GB` on disk and contains only the integer-labelled mask hypotheses used for downstream analysis and tracking. The auxiliary Omnipose distance and flow outputs were omitted from this release to keep the archive practical to distribute.

## What is in this repository

The main array is:

- `data(hypothesis, trench, time, y, x)` with shape `(17, 2795, 721, 164, 34)`

and:

- dtype: `uint8`
- chunks: `(1, 1, 721, 164, 34)`
- axis names: `["hypothesis", "trench", "time", "y", "x"]`

Each `data[h, trench, time]` slice is a single instance-labelled segmentation mask for one trench at one timepoint. Background is labelled `0`; cell instances are stored as positive integer labels.

The `hypothesis` axis corresponds to Omnipose inference at different saved training checkpoints:

- `[50, 300, 550, 800, 1050, 1300, 1550, 1800, 2050, 2300, 2550, 2800, 3050, 3300, 3550, 3800, 3999]`

These checkpoint values are stored in:

- `data.attrs["hypothesis_epoch_values"]`

## Metadata included in the store

Although this is a masks-only release, the store includes enough metadata to map masks back to the experiment and trench geometry. Root attributes include:

- `source_nd2`
- `source_zarr`
- `experiment_name`
- `pixel_size_um`
- `n_trenches`, `n_times`, `trench_height`, `trench_width`
- `registration_params`
- `source_metadata_version`
- `source_acquisition_metadata`
- `source_subset_metadata`
- `segmentation_hypothesis_epochs`
- `segmentation_model_checkpoint_template`
- `segmentation_run_stem`
- `segmentation_params`

The store also includes a full `trench_mapping` table in the root attributes, giving for each trench:

- `trench_id`
- source `fov`
- `lane_index`
- crop bounds `x_left`, `x_right`, `y_top`, `y_bottom`
- `orientation`
- `needs_flip`

This means the mask archive can be interpreted on its own for segmentation benchmarking and hypothesis-aware tracking, although users who need the underlying images should pair it with the companion trench image store.

## Experimental context

The masks correspond to extracted trench movies of *E. coli* K-12 MG1655 cells in a mother-machine experiment monitoring dormancy exit. Cells were held under spent LB plus pluronic for the first `60 min` and then switched to fresh LB plus pluronic to induce resuscitation. Imaging was performed at approximately `30 s` intervals for `721` frames.

The original image data can be found at https://huggingface.co/datasets/ghardo/scientific_data_2026_images

## How the masks were generated

Synthetic training data were generated with SyMBac, and an Omnipose model was trained on the resulting mother-machine images. To preserve segmentation uncertainty, inference was run at multiple saved checkpoints rather than exporting only a single final mask set.

The released masks therefore represent `17` alternative segmentation hypotheses for the same underlying trench movies. In downstream use, these hypotheses can be:

- compared to quantify segmentation stability across checkpoints
- used as input to multi-hypothesis or repair-based tracking workflows
- reduced to a single preferred hypothesis if only one segmentation per frame is needed

In the associated tracking workflow, the final-checkpoint masks are used for an initial Trackastra pass, and ambiguous local intervals are then revisited using the full hypothesis set.

## Data format and access

This is a directory-style Zarr store and is best accessed with `zarr`, `dask.array`, or other Zarr-aware tools.

### Minimal `zarr` example

```python
import zarr

root = zarr.open(
    "20260307_SB7_exit_snake_V4_1.segmentation_masks_multi_epoch_uint8_masks_only.zarr",
    mode="r",
)

masks = root["data"]  # (hypothesis, trench, time, y, x)
epochs = masks.attrs["hypothesis_epoch_values"]
pixel_size_um = root.attrs["pixel_size_um"]
```

### `dask.array` example

```python
import dask.array as da

masks = da.from_zarr(
    "20260307_SB7_exit_snake_V4_1.segmentation_masks_multi_epoch_uint8_masks_only.zarr/data"
)

print(masks.shape)   # (17, 2795, 721, 164, 34)
print(masks.dtype)   # uint8
```

### Selecting one hypothesis

```python
import zarr

root = zarr.open(
    "20260307_SB7_exit_snake_V4_1.segmentation_masks_multi_epoch_uint8_masks_only.zarr",
    mode="r",
)

masks = root["data"]
epochs = masks.attrs["hypothesis_epoch_values"]

final_checkpoint_masks = masks[-1]   # epoch 3999
early_checkpoint_masks = masks[0]    # epoch 50
```

## Recommended uses

This dataset is well suited to:

- benchmarking bacterial instance segmentation under low-contrast mother-machine imaging
- testing multi-hypothesis tracking methods
- measuring checkpoint-to-checkpoint segmentation agreement
- extracting morphology from alternative mask hypotheses
- studying how segmentation uncertainty propagates into downstream lineage reconstruction

## Limitations

- This repository contains masks only, not the trench image data.
- The release omits Omnipose distance and flow fields.
- The masks represent agreement across checkpoints of one training pipeline, not independent human ground truth.
- Labels are stored as `uint8`, so this format assumes the number of instances per trench frame fits within that range.
- The store is not packaged for the Hugging Face `datasets` tabular loader; use Zarr-aware tools directly.

## Companion data

For the underlying image data and full acquisition metadata, use the companion image repository (https://huggingface.co/datasets/ghardo/scientific_data_2026_images) containing:

- `20260307_SB7_exit_snake_V4_1_with_metadata.trenches.zarr`

Together, the image store and the mask-hypothesis store support segmentation benchmarking, morphology extraction, and lineage reconstruction workflows.


## Repository structure

```text
20260307_SB7_exit_snake_V4_1.segmentation_masks_multi_epoch_uint8_masks_only.zarr/
├── zarr.json
└── data/
    └── zarr.json and chunk data
```