napari-tmidas

Need fast batch processing for confocal & whole-slide microscopy images of biological cells and tissues?

This open-source napari plugin integrates state-of-the-art AI + analysis tools in an interactive GUI with side-by-side result comparison! Transform, analyze, and quantify microscopy data at scale including deep learning - from file conversion to segmentation, tracking, and analysis.

✨ Key Features

🤖 AI Methods Built-In

• Virtual staining (VisCy) • Denoising (CAREamics) • Spot detection (Spotiflow) • Segmentation (Cellpose, Convpaint) • Tracking (Trackastra, HOCT, Ultrack)
• Auto-install in isolated environments • No dependency conflicts • GPU acceleration

🔄 Universal File Conversion

• Convert LIF, ND2, CZI, NDPI, Acquifer → TIFF or OME-Zarr
• Preserve spatial metadata automatically

⚡ Batch Processing

• Process entire folders with one click • 40+ processing functions • Progress tracking & quality control
• One Dimension Order setting, shared by every function — set it once per batch, not per function

� Interactive Workflow

• Side-by-side table view of original and processed images • Click to instantly compare results • Quickly iterate parameter values • Real-time visual feedback

�📊 Complete Analysis Pipeline

• Segmentation → Tracking → Quantification → Colocalization

🚀 Quick Start

Supports Python 3.11+; commands below use Python 3.12.

# Install napari and the plugin
mamba create -y -n napari-tmidas -c conda-forge python=3.12
mamba activate napari-tmidas
pip install "napari[all]"
pip install napari-tmidas

# Launch napari
napari

Then find napari-tmidas in the Plugins menu. Watch video tutorials →

💡 Tip: AI methods (SAM2, Cellpose, Spotiflow, etc.) auto-install into isolated environments on first use - no manual setup required!

⚠️ Before a batch run: set Dimension Order (top of the batch widget) to match your data — TZYX for a 3D time series, ZYX for a Z-stack, TYX for a 2D movie. Most microscopy TIFFs carry no usable axis metadata, so on Auto a function that builds 3D objects cannot tell Z from T: it will either stop with an error or label the same object once per Z slice. The widget reads each file's rank up front, shows it next to the dropdown, and warns before the run starts.

📖 Documentation

AI-Powered Methods

Core Workflows

• File Conversion - Multi-format microscopy file conversion (LIF, ND2, CZI, NDPI, Acquifer)
• Batch Processing - All 40+ processing functions in one place
• Frame Removal - Interactive human-in-the-loop frame removal from time series
• Label-Based Cropping - Interactive ROI extraction with label expansion
• Quality Control - Visual QC with grid overlay
• Quantification - Extract measurements from labels

Advanced Features

• Batch Crop Anything - Interactive object cropping with SAM2
• Batch Label Inspection - Manual label verification and editing, with one-click delete/relabel across all timepoints, click-to-split for merged objects, and click-to-merge-neighbors for over-segmented ones
• Multichannel Processing - Channel selection and per-channel processing

💻 Installation

Step 1: Install napari

mamba create -y -n napari-tmidas -c conda-forge python=3.12
mamba activate napari-tmidas
python -m pip install "napari[all]"

Step 2: Install napari-tmidas

🖼️ Screenshots

File Conversion Widget

Convert proprietary formats to open standards with metadata preservation.

Batch Processing Interface

Select files → Choose processing function → Run on entire dataset.

Label Inspection

Inspect and manually correct segmentation results.

SAM2 Crop Anything

Interactive object selection and cropping with SAM2.

📋 TODO

Memory-Efficient Streaming

Most of this is done. Batch processing no longer materializes whole stacks: the worker keeps large inputs lazy and streams results back to disk block by block (256 MB budget), 15 functions map their existing body over blocks via the @chunked decorator, and 6 more own their I/O outright via skip_load. Measured end to end on a real (31, 2, 57, 2720, 2720) uint16 acquisition — 52 GB dense — Gamma Correction peaks at 3.15 GB RSS in 8.8 min, byte-identical to the dense path. Convpaint prediction, Cellpose segmentation and Trackastra tracking all write per-timepoint now. The mechanism is documented in _chunked.py, and the behaviour is pinned by TestZarrOutputStreaming, TestSplitChannelsStreaming, TestLazyTiffLoading, TestCLAHEDaskMemory and TestRollingBallPerPlane.

What is left:

• CAREamics denoising and VisCy virtual staining still run dense. Both allocate the full output array and take the input as a NumPy array. Neither has been audited for real — that needs their dedicated virtualenvs installed.
• ~14 registered functions still scale linearly with input size. That is now a known list rather than an unknown one. Each needs either a @chunked conversion or a check that it cannot take one: functions using global statistics (Convert to 8-bit rescales by the whole-stack min/max) and functions with cross-block topology (Mirror Labels) both resist it.
• The structural decision. Dense functions opt into laziness one at a time, by accepting _source_filepath. Whether to keep converting them individually or change the worker's contract for all of them at once is still open.

Other Known Issues

• Resize Zarr by YX Scale (OME-Zarr native) writes zarr v3 stores: attributes live in zarr.json (multiscales nested under an ome key), not in a v2 .zattrs. Anything downstream that reads .zattrs directly will find no file there — use ome_output_utils._read_root_attrs(), which handles both layouts.

🤝 Contributing

Contributions are welcome! Please ensure tests pass before submitting PRs:

pip install tox
tox

📄 License

BSD-3 License - see LICENSE for details.

🐛 Issues

Found a bug or have a feature request? Open an issue

🙏 Acknowledgments

Built with napari and powered by:

AI/ML Methods:

• Cellpose • Convpaint • VisCy • CAREamics • Spotiflow • Trackastra • HOCT • Ultrack • SAM2

Core Scientific Stack:

• NumPy • scikit-image • PyTorch

File Format Support:

• OME-Zarr • tifffile • nd2 • pylibCZIrw • readlif