napari-vipp 0.7.0a1

Visual image processing pipelines for napari

napari-vipp

Alpha Disclaimer

napari-vipp is an early alpha build in active development.

• Expect breaking changes between releases.
• Workflows, node parameters, and file compatibility may change.
• Validate outputs before scientific interpretation or publication use.

napari-vipp is an early prototype of a napari-native Visual Image Processing Pipeline (VIPP): an interactive node-graph workflow composer for bioimage analysis.

The project is exploring a workflow where the graph is the main work surface: users add processing nodes, connect outputs to inputs, tune parameters, inspect stage outputs in napari, and compare mask overlays or processing branches while the pipeline updates live.

This is research/prototype software. The current code is useful for exploring interaction patterns and metadata-aware pipeline execution, but it is not yet a released production analysis package.

Acknowledgement

If VIPP contributes to your work, please acknowledge the project by name ("napari-vipp") and include a link to the repository: https://github.com/rensutheart/napari-vipp

Formal citation metadata can be added in a future release once a citation target (for example DOI or manuscript) is available.

Current Status

The prototype currently supports:

• a napari npe2 plugin manifest and dock widget;
• a large pan/zoom Qt graph canvas;
• toolbar graph zoom controls with a 100% reset plus Ctrl/trackpad wheel zoom;
• draggable node cards with input/output ports and curved connectors;
• adding nodes from a categorized, fuzzy-searchable node library;
• connecting nodes by dragging or click-to-connect from output to input ports;
• explicit image source nodes for napari layers, files, or bundled samples;
• quick selected-output saving plus graph-level save nodes;
• per-node thumbnails with global show/hide, Slice/MIP preview modes, contrast modes (Percentile, Min-max, Raw), and monochrome colormaps;
• optional per-node thumbnail disabling for heavier workflows;
• selected-node parameter controls in the inspector;
• slider plus numeric entry controls with soft range expansion where useful;
• right-panel output histograms plus cutoff-node input histograms with slice/stack and linear/log modes;
• compact node metadata plus detailed selected-node metadata;
• normalized axes, channel, acquisition, source, and provenance metadata;
• OME-TIFF, ImageJ TIFF, conventional TIFF, OME-Zarr 0.4/0.5, and common raster image import plus 2D raster export;
• adaptive image/series selection for multi-image sources;
• table outputs for object, intensity, skeleton, merged, and annotated results;
• image/mask/label pinning as persistent napari preview layers;
• generated inspect layers for full-resolution napari review.

Image Metadata

The graph carries an explicit image state object alongside each node output. That state records:

• shape and dtype;
• axis names and axis types, such as t, c, z, y, x;
• units, scale, and origin/translation where available;
• value range, bit depth, memory estimate, and binary-value hints;
• source layer and operation history.
• channel names, fluorophore/wavelength fields where available;
• acquisition and stable source identity records.

When a napari layer provides OME-NGFF-style multiscales metadata, VIPP reads the axis definitions and coordinate transforms. When the source is a plain array without reliable metadata, VIPP falls back to inferred axes and labels that fallback explicitly.

OME-TIFF metadata and local OME-Zarr 0.4/0.5 images are supported through the shared headless I/O layer. OME-Zarr label groups, HCS plate browsing, generated pyramids, remote stores, and full operation-level lazy execution remain future work. See docs/io-user-guide.md for the current format contract.

Sample Data

The plugin contributes synthetic fluorescence-like sample data:

• VIPP synthetic volume: grayscale ZYX stack;
• VIPP synthetic multichannel volume: CZYX volume with three probe-like channels;
• VIPP synthetic time-lapse multichannel: TCZYX time-lapse, multichannel stack.

The multichannel samples use separate intensity channels, not baked RGB images. Graph thumbnails render these as fluorescence-style pseudo-color composites while preserving the underlying channel axis in the carried metadata. When the full sample suite is open, the workflow automatically starts from the VIPP synthetic time-lapse multichannel layer so the input metadata should read as TCZYX. The simpler grayscale and CZYX examples are still available in the toolbar input selector and in the graph-level Image Source node.

Open sample data from napari:

File > Open Sample > VIPP synthetic microscopy samples

Node Library

The current node catalogue includes:

• Image Data:
  • Source & Output:
    • Image Source
    • Save Image
  • Axes & Regions:
    • Crop Stack
    • Select Axis Slice
    • Reorder Axes
    • Set Pixel Size / Units
    • Rescale Axes
  • Channels & Composites:
    • Extract Channel
    • Combine Channels
    • Split Channels
    • Composite → RGB
  • Utilities:
    • Convert Dtype
  • Math & Logic:
    • Calculate New Image
    • Add
    • Subtract
    • Ratio
    • Mask Image
    • Logical AND
    • Logical OR
    • Logical XOR
    • Invert
• Intensity & Contrast:
  • Linear Scale + Offset
  • Gamma Correction
  • Rescale Intensity
  • Normalize
  • Clip
• Filtering:
  • Smoothing & Denoising:
    • Average Blur
    • Gaussian Blur
    • Gaussian Blur 3D
    • Median Filter
    • Bilateral Filtering
    • Non-Local Means
  • Edge & Detail:
    • Difference of Gaussians
    • Unsharp Mask
    • Sobel Edges
    • Canny Edges
    • Laplace Filter
• Projection:
  • Maximum Projection
  • Project Image (axis-aware dropdown plus multiple projection methods)
  • Orthogonal Projection
• Segmentation:
  • Global Thresholds:
    • Otsu Threshold
    • Triangle Threshold
    • Li Threshold
    • Yen Threshold
    • Isodata Threshold
    • Minimum Threshold
    • Binary Threshold
    • Hysteresis Threshold
  • Local Thresholds:
    • Adaptive Mean Threshold
    • Adaptive Gaussian Threshold
    • Sauvola Threshold
    • Niblack Threshold
• Morphology:
  • Dilation
  • Erosion
  • Opening
  • Closing
  • Top Hat
  • Black Hat
  • Morphological Gradient
  • Fill Holes
  • Remove Small Objects
  • Skeletonize
• Label Operations:
  • Label Connected Components
  • Filter Labels By Volume
  • Filter Labels By Property
  • Clear Border Objects
  • Relabel Sequential
• Measurements:
  • Measure Objects
  • Measure Objects + Intensity
  • Analyze Skeleton
  • Merge Tables
  • Select Table Columns
  • Add Metadata Columns

Histogram-based automatic threshold nodes show Threshold uses on stack inputs. Stack histogram computes one cutoff from the whole grayscale input and applies it to the full image; Slice histogram computes a separate cutoff per displayed plane while still producing a full-stack mask. The control is hidden for 2D inputs. These nodes also show the input histogram used for threshold selection with a live marker at the chosen threshold.

The label pipeline converts binary masks into integer object IDs. Connected components can run over full ZYX volumes or independently over YX images. Volume filtering currently uses pixel/voxel counts, preserves retained IDs, and leaves compact renumbering to the explicit Relabel Sequential node. Its volume sliders use the largest observed object as their data-aware upper bound and a logarithmic scale for useful control across small and large structures; the numeric fields still accept exact values up to one billion. Selecting Filter Labels By Volume also shows the incoming object-volume distribution above the regular histogram. Dashed minimum and enabled maximum markers update with the filter controls. Its Log volume axis toggle is enabled by default and can be switched off for a linear distribution.

Filter Labels By Property accepts named Labels and Measurements table inputs and keeps or removes labels using any numeric table column, including area/volume, intensity, and skeleton/network measurements. It preserves label IDs and leaves compact renumbering to Relabel Sequential.

Clear Border Objects accepts either a binary mask or integer labels and preserves that semantic type. In 3D it can remove objects touching all ZYX volume boundaries or only the lateral YX boundaries, with an optional border buffer. Timepoints and channels are processed independently.

Fill Holes accepts binary masks and defaults to metadata-aware processing: 2D images are filled in YX, while true z-stacks are filled as complete ZYX volumes. An advanced 2D-per-slice mode remains available for deliberately slice-wise segmentations. A maximum size of 0 fills every enclosed hole; positive values fill only holes up to the selected pixel area or voxel volume. The inspector hides 3D for true 2D inputs and warns when slice-wise filling is selected for a z-stack.

Remove Small Objects accepts binary masks or integer labels and preserves the connected semantic type. It removes objects below a minimum pixel area or voxel volume, supports 2D or 3D processing, and exposes connectivity for mask inputs. Its logarithmic size control is bounded by the largest observed input object. For labels-only cleanup with both minimum and maximum cutoffs, use Filter Labels By Volume.

Measure Objects accepts a label image and produces a table output instead of an image. The default measurement set includes label ID, pixel/voxel area or volume, calibrated physical area or volume when spatial scale metadata is available, centroid, bounding box, equivalent diameter, extent, and Euler number. Optional checkboxes add shape descriptors, axis/inertia descriptors, and 2D boundary descriptors. The 2D boundary group is hidden for true 3D inputs because perimeter, Crofton perimeter, orientation, and eccentricity are 2D concepts in the current implementation. Table outputs show a row preview in the inspector and can be saved as CSV or TSV.

Measure Objects + Intensity is the first named multi-input measurement node. It has separate Labels and Intensity image input ports, then outputs the basic object morphology columns plus per-label mean, minimum, maximum, sum, and standard deviation intensity. It exposes the same optional morphology groups as Measure Objects. The example workflow examples/red-channel-object-intensity-measurements.json demonstrates this pattern.

Merge Tables joins two or more table outputs into a single table. In auto mode it joins on stable identity columns such as t_index and label_id; when no identity columns are shared, equal-length tables can be joined by row position. Select Table Columns keeps, drops, or reorders table columns while preserving row order and column units. Add Metadata Columns appends constant treatment, replicate, batch, or condition columns before CSV/TSV export. The example workflow examples/red-channel-merged-measurement-table.json demonstrates a PCA-oriented table assembly path.

Skeletonize accepts a binary mask and produces a skeleton mask using metadata-aware 2D or 3D processing. Analyze Skeleton accepts a skeleton mask and outputs a per-component table with skeleton voxel count, endpoint voxels, junction voxels, isolated nodes, branch/graph edge counts, voxel-graph edge count, cycle count, per-block component count, component voxel fraction, and skeleton length in pixel/voxel and physical units when scale metadata is available. These nodes are generic and are intended for mitochondria, neurites, vessels, fibers, hyphae, and other curvilinear structures.

Extract Channel pulls one selected channel from a multichannel image. Split Channels is its bulk counterpart: it emits one output port per channel in the image (losslessly, preserving dtype), with the port count following the true channel count. Each channel also preserves the semantic type of its input, so splitting a threshold mask produces mask ports that connect directly to Label Connected Components. Combine Channels is the inverse multi-input node: set the expected channel/input count, connect that many upstream images, and it stacks them into an explicit multichannel output. Channel pseudo-colours are carried as metadata from OME sources, Image Source overrides, Combine Channels, or the Assign Channel Colors pass-through node. Composite → RGB maps a multichannel composite to a channel-last RGB image. Auto mode preserves true RGB/RGBA inputs, and otherwise blends all channels by their carried pseudo-colours, so yellow contributes to red and green and cyan contributes to green and blue. Manual red/green/blue selectors remain available for forced single-channel plane mapping. Calculate New Image is a multi-input image-math node that applies comma separated weights to connected inputs and then adds an offset.

Mask Image is a typed two-input node with separate Image and Mask ports. The mask port accepts binary masks or labels, treating nonzero values as inside the mask. Spatial masks can be broadcast over compatible image axes, including channel-last RGB/RGBA images and common channel-first multichannel arrays, so a YX or ZYX mask can mask all colour/channel planes without first splitting the image.

Image Source can point to an existing napari layer, a local .npy, TIFF, OME-Zarr, or common raster source such as PNG/JPEG/BMP/GIF/WebP, or one of the bundled synthetic samples. Set Pixel Size / Units repairs missing or incorrect input calibration by setting X/Y pixel size, optional Z step size, and the shared physical unit carried in downstream metadata. Scale-aware nodes such as Orthogonal Projection use this metadata to preserve physical proportions for anisotropic z-stacks. Rescale Axes changes the sampled pixel grid along X/Y/Z with optional X/Y aspect-ratio locking, nearest-neighbor through spline interpolation choices, and anti-aliasing for intensity-image downsampling; it updates physical scale metadata inversely to the requested scale factors. Save Image passes data through unchanged and, when Auto-save on update is set to on, writes the node input to disk every time the graph recomputes. For quick interactive work, the inspector also provides Save selected output... for the currently selected node; that dialog defaults to TIFF but also allows .npy and PNG/JPEG-style formats when the selected output is 2D. TIFF output is written in ImageJ hyperstack format when axis metadata is available, and binary masks are saved as 8-bit 0/255 values.

Development

Create a local environment and install in editable mode:

python -m venv .venv
.venv\Scripts\activate
python -m pip install -e ".[dev]"

Validate and test:

python -m npe2 validate src/napari_vipp/napari.yaml
python -m ruff check .
python -m pytest

Launch napari manually and open the widget from:

Plugins > VIPP Workflow (napari-vipp)

Or launch the local sample app:

python scripts\launch_vipp_sample.py

To start directly from the multichannel Otsu-to-label workflow:

python scripts\launch_vipp_label_workflow.py

The same graph can be loaded manually from examples/otsu-red-channel-labels.json. A second review workflow, examples/red-channel-object-intensity-measurements.json, demonstrates the named Labels plus Intensity image input slots on Measure Objects + Intensity.

Roadmap

Near-term development priorities:

• axis-aware channel selectors that show probe names instead of only numbers;
• grouped table summaries by condition/time/source;
• skeleton QC feature masks, branch labels, and short-branch pruning;
• distance transforms, marker generation, and marker-controlled watershed;
• richer 3D mesh morphology and calibrated physical variants for extended length/shape measurements;
• fluorescence background correction;
• OME-Zarr pyramids, label colors/properties, and preview-resolution selection;
• plate/well/field browsing, remote reads, batch execution, and memory-aware lazy execution;
• richer non-image outputs, including points and scalar summaries;
• batch execution over files, positions, channels, z-slices, and timepoints.

See docs/planning.md for broader planning, docs/io-user-guide.md for current I/O behavior, docs/ome-io-plan.md for the accepted OME architecture, and docs/research-and-publication.md for the evidence and publication record.