Datasets:
Fish-X LC-NE Test
Fish-X LC-NE Test provides zebrafish electron microscopy images and corresponding reference labels for SAEM²-SAEvM³ and its interactive Neuroglancer demo. It is the repository's supplied test dataset.
Setup tutorial · Repository and installation · Window browser · Nucleus segmentation
Three demo options
| Option | Data needed | Views |
|---|---|---|
1. examples (default) |
About 114 MiB already bundled in the SAvEM3 repository; no dataset download or login | Full lc labels/meshes and five bounding boxes, followed by five process windows |
2. full |
Download and extract the full dataset, about 201 GB | Full raw/label overview with optional context layers, followed by the same five process windows |
3. nuclei |
Extracted raw overview at 32 x 32 x 30 nm | Positive-center SAM2 nucleus segmentation on the overview, plus the same five process windows |
The five process windows cover neurons 30001-30005. Each contains 512 x 512 x 63 voxels at 8 x 8 x 30 nm, with raw images, instance labels, a binary reference mask, and a supplied local realignment field. Local realignment is enabled by default in every process window.
Fresh process reviews load eight saved tracing objects and 55 annotation
points from the repository's bundled starter session. Existing reviews are
preserved. Add --no-starter-session with a new output directory to start blank.
A fresh nucleus review starts without saved nuclei.
Launch the demo
Install Probe-EM and SAM2 and obtain a matching SAM2 checkpoint. Run the commands below from the SAvEM3 repository root. Model weights are separate from the dataset. Replace example paths with your local paths and keep the Python server running while using the viewer.
1. Bundled lc overview and five process windows
python probe-em/scripts/demo_windows.py \
--dataset examples --checkpoint /models/sam2_hiera_tiny.pt \
--output-dir .cache/fish-x-review --open
The launcher verifies, extracts, and caches the bundled examples automatically.
The Fish-X LC-NE Test overview shows lc and test_windows. Raw images
are available inside each process window. This option works without a Hugging
Face account.
2. Full overview and the same five windows
This public dataset uses manually approved access requests. Request access on
the dataset page and authenticate with the approved account before downloading.
Follow the setup tutorial to restore all seven zfishlc_* volumes
under a common data directory. Allow about 201 GB for downloads and another
201 GB for extraction, plus review outputs and caches.
python probe-em/scripts/demo_windows.py \
--dataset full --data-root /data/Fish-X-LC-NE-Test \
--checkpoint /models/sam2_hiera_tiny.pt \
--output-dir .cache/fish-x-review --open
The overview adds whole-volume raw images. Mito, syn, ves, lcpre, and lcsyn are
optional display layers available only in this overview and start hidden.
The launcher uses the five examples already bundled in the repository; no
separate extraction of demo/lc_ne_process_windows.zip is needed.
3. Nucleus segmentation on the 32 nm overview
python probe-em/scripts/demo_windows.py \
--dataset nuclei --data-root /data/Fish-X-LC-NE-Test \
--checkpoint /models/sam2_hiera_tiny.pt \
--output-dir .cache/fish-x-review --open
This mode needs zfishlc_raw/info and the complete 32_32_30 raw scale under
--data-root. Coarser raw scales are used when available. Other full-volume
datasets are not required. The lc labels and window boxes are bundled and start
hidden; the five process examples remain available in the selector.
Use P → V → C: place one positive point at a nucleus center, run SAM2
propagation above and below that section, then confirm the nucleus. Each accepted
nucleus keeps a separate colored mask, and the next nucleus receives a new color.
All active and saved centers stay in one nucleus_centers annotation layer.
Negative points are not used.
Inference always reads 32 x 32 x 30 nm raw data in bounded blocks, regardless of display zoom, without additional local realignment. Propagation stops on empty sections or reports a partial result when a crop or section limit is reached. Inspect the preview before confirming. See the nucleus guide for limits and saved files.
Review controls
Use the top-bar selector or Previous / Next to switch views. Views open with
the 4panel-alt layout, visible axes, and showSlices=false. Shortcut labels are
uppercase; press these letter keys without Shift.
| Control | Action |
|---|---|
| D / F | Previous / next section, including while dragging |
| P | Add a positive process point, or place the active nucleus center |
| N | Add a negative point in a process window |
| I | Predict the current section in a process window |
| B / E | Set the process-window propagation start / end |
| V | Propagate through the selected process range, or automatically above and below a nucleus center |
| C | Save the current object and its points, then start the next in a new color |
| Y, then Enter, Enter | Accept/save the combined mask after two confirmations; Esc cancels |
| R | Toggle local realignment in a process window |
| X | Cancel a running task |
Accepted process objects retain their positive and nonempty negative-point
layers. Nuclei retain their centers in the single shared layer. Final acceptance
keeps the individual masks and points available alongside the combined mask.
ground_truth uses neutral gray, distinct from the first five object colors.
See the interactive guide
for process refinement and multi-object review.
Reuse the output directory to restore accepted objects. Process results are
stored in OUTPUT/neuron_<ID>/interactive/; nucleus results are stored in
OUTPUT/nuclei/interactive/. Review outputs are separate from the source data.
Contents
| Dataset | Files | Payload GiB | ZIP parts |
|---|---|---|---|
| zfishlc_raw | 19,302 | 132.304 | 27 |
| zfishlc_mito | 19,302 | 17.658 | 4 |
| zfishlc_syn | 19,302 | 17.584 | 4 |
| zfishlc_ves | 19,302 | 19.761 | 4 |
| zfishlc_lc | 19,456 | 0.044 | 1 |
| zfishlc_lcpre | 19,529 | 0.031 | 1 |
| zfishlc_lcsyn | 21,005 | 0.014 | 1 |
All seven volumes provide six resolutions in XYZ nanometers: 8 x 8 x 30,
16 x 16 x 30, 32 x 32 x 30, 64 x 64 x 60, 128 x 128 x 120, and
256 x 256 x 240. The finest grid has shape 8640 x 8160 x 1583 and offset
[192, 288, 0].
Each ZIP64 part is an independent archive; do not concatenate parts. The included
restore.py verifies archive and member checksums during extraction. The full
volumes contain legacy mesh filenames with colons; use a WSL Linux filesystem
for the full option on Windows. The bundled examples, including their lc
overview, support native Windows, macOS, and Linux.
Reference labels cover selected structures, not every neuron or nucleus in the volume. Mito, syn, and ves are grayscale context images. Interactive predictions and user annotations are saved separately by the demo.
License
The dataset is licensed under Apache-2.0, matching the SAvEM3 repository. See NOTICE for included software notices.
- Downloads last month
- 27