Skip to content

advanced_processing.md

Latest commit

 

History

History
408 lines (274 loc) · 11.5 KB

advanced_processing.md

File metadata and controls

408 lines (274 loc) · 11.5 KB

CAREamics Denoising

Overview

Image denoising using CAREamics (Content-Aware Image Restoration). This processing function removes noise from microscopy images using deep learning-based methods, including Noise2Void (N2V) and CARE models.

Features

  • Multiple Model Support: Works with Noise2Void, CARE, and custom trained models
  • 2D and 3D Data: Handles both 2D and 3D microscopy images
  • Tile-Based Processing: Efficiently processes large images using tiling
  • Test-Time Augmentation: Optional TTA for improved results
  • Automatic Environment Management: Creates dedicated environment if needed

Installation

CAREamics will automatically install into a dedicated careamics-env environment when first used.

Parameters

checkpoint_path (string, required)

Path to the CAREamics model checkpoint file (.ckpt).

How to obtain:

  • Train your own model using CAREamics
  • Download pre-trained models from CAREamics model zoo
  • Use models provided by your lab/collaborators

tile_size_x (int, default: 32, range: 16-512)

Tile size in X dimension for processing.

tile_size_y (int, default: 32, range: 16-512)

Tile size in Y dimension for processing.

tile_size_z (int, default: 0, range: 0-256)

Tile size in Z dimension (for 3D data). Set to 0 for 2D images.

batch_size (int, default: 1, range: 1-16)

Number of tiles to process simultaneously.

  • Increase for faster processing (requires more memory)
  • Decrease if you encounter out-of-memory errors

use_tta (bool, default: True)

Use test-time augmentation for better denoising results.

  • True: Better quality but slower
  • False: Faster but potentially lower quality

force_dedicated_env (bool, default: False)

Force using dedicated environment even if CAREamics is available in the main environment.

Usage

Prerequisites

  1. Trained CAREamics model: You need a checkpoint file (.ckpt)
  2. Compatible image data: Must match the dimensions the model was trained on

In napari-tmidas

  1. Open Plugins > T-MIDAS > Image Processing
  2. Browse to your folder containing noisy images
  3. Select "CAREamics Denoise (N2V/CARE)"
  4. Configure parameters:
    • Set checkpoint_path to your model file
    • Adjust tile sizes based on your image dimensions
    • Enable use_tta for better results
  5. Click "Start Batch Processing"

Output

Suffix: _denoised

Produces denoised images with:

  • Same dimensions as input
  • Same data type as input
  • Reduced noise while preserving structures

Tips

  1. Model Selection: Use models trained on similar data types
  2. Tile Sizes: Match the receptive field of your model
  3. TTA: Use for final results, disable for quick tests
  4. Memory: Reduce batch_size and tile sizes if out of memory

Credits

Spotiflow Spot Detection

Overview

Accurate spot detection for fluorescence microscopy using Spotiflow, a deep learning-based method designed specifically for detecting fluorescent spots and puncta.

Features

  • Pre-trained Models: Multiple models for different microscopy types
  • 2D and 3D Support: Handles both 2D and 3D image stacks
  • High Accuracy: Optimized for sub-pixel spot localization
  • Label Output: Can generate label masks from detected spots
  • Automatic Environment Management: Dedicated environment created if needed

Installation

Spotiflow will automatically install into a dedicated spotiflow-env environment when first used.

Pre-trained Models

  • general: General-purpose model for various spot types
  • smfish: Optimized for smFISH (single-molecule FISH)
  • spots_3d: For 3D volumetric spot detection
  • Custom models: Provide path to your trained model

Parameters

pretrained_model (string, default: "general")

Name of the pre-trained Spotiflow model or path to custom model.

model_path (string, default: "")

Path to custom trained Spotiflow model (overrides pretrained_model).

prob_thresh (float, default: 0.4, range: 0.0-1.0)

Probability threshold for spot detection.

  • Lower: More spots detected (may include false positives)
  • Higher: Fewer, higher-confidence spots

spot_radius (int, default: 3, range: 1-20)

Radius for creating label masks around detected spots (in pixels).

force_cpu (bool, default: False)

Force CPU processing even if GPU is available.

generate_labels (bool, default: True)

Generate label masks from detected spots (vs. just coordinates).

Usage

In napari-tmidas

  1. Open Plugins > T-MIDAS > Image Processing
  2. Browse to folder with fluorescence images
  3. Select "Spotiflow Spot Detection"
  4. Configure parameters based on your data
  5. Click "Start Batch Processing"

Output

  • With labels: Label image with each spot as unique region
  • Without labels: Returns detected spot coordinates

Tips

  1. Model Selection: Choose model matching your imaging modality
  2. Threshold Adjustment: Lower threshold for dim spots, raise for bright spots
  3. 3D Data: Use spots_3d model for volumetric data
  4. GPU: Enable GPU for faster processing of large datasets

Credits

SciPy Filters

Overview

Image processing functions using SciPy's ndimage module for filtering and morphological operations.

Functions

Resize Labels (Nearest, SciPy)

Suffix: _scaled

Resize label images while preserving label integrity using nearest-neighbor interpolation.

Parameters:

  • scale_factor (float, default: 1.0, range: 0.01-10.0): Scaling factor
    • < 1.0: Shrink labels

    • 1.0: Enlarge labels

    • 1.0: No change

Use case: Scale segmentation masks to match differently sized images.

Technical details:

  • Uses scipy.ndimage.zoom with grid_mode=True
  • Preserves label values exactly (no interpolation)
  • Centers resized objects in original array dimensions
  • Maintains spatial relationships

Subdivide Labels into 3 Layers

Suffix: _layers

Subdivide each labeled object into 3 concentric layers (outer, middle, inner).

Parameters:

  • is_half_body (bool, default: False): Enable for objects cut in half

Use case: Analyze cell compartments or tissue layers separately.

Output: Single label image where each layer gets unique ID:

  • Original object ID=1 → outer layer=1, middle layer=1001, inner layer=2001
  • Original object ID=2 → outer layer=2, middle layer=1002, inner layer=2002

Technical details:

  • Uses distance transform for layer calculation
  • Layers are approximately equal volume
  • Half-body mode creates layers as if object were complete

Installation

pip install scipy

Scikit-Image Filters

Overview

Image enhancement and filtering functions using scikit-image library.

Functions

Resize Image by YX Scale (skimage)

Suffix: _yx_resized

Resize intensity images by a single YX scale factor while preserving non-spatial axes (T/C/Z).

Parameters:

  • scale_factor (float, default: 0.5): YX scaling factor
    • 0.5: Half-size in Y and X (about 4× fewer pixels)
    • 1.0: No resize
    • >1.0: Upscale

Use case: Speed up downstream segmentation while keeping proportional geometry.

Resize Zarr by YX Scale (OME-Zarr native)

Suffix: _yx_resized

Resizes zarr inputs using lazy OME-Zarr/dask I/O and writes output zarr directly.

Parameters:

  • scale_factor (float, default: 0.5): YX scaling factor

Technical details:

  • Preserves source pyramid depth
  • Preserves axes metadata and updates coordinate transforms after resize
  • Supports channel selection

CLAHE (Adaptive Histogram Equalization)

Suffix: _clahe

Apply Contrast Limited Adaptive Histogram Equalization to enhance local contrast.

Parameters:

  • clip_limit (float, default: 0.01): Contrast clipping limit (0.001-0.1)
    • Higher: More contrast, may amplify noise
    • Lower: Subtler enhancement
  • kernel_size (int, default: 0): Local region size (0=auto)
    • Smaller: Enhance small features
    • Larger: Enhance large features

Use case: Enhance weak bright features in dark images (membranes, fine structures).

Technical details:

  • Works locally, prevents over-brightening of background
  • Auto-calculates kernel size if not specified
  • Much better than global histogram equalization for microscopy

Additional scikit-image Filters

Other available filters include:

  • Gaussian Blur: Smooth images
  • Median Filter: Remove salt-and-pepper noise
  • Edge Detection: Sobel, Canny edge detectors
  • Morphological Operations: Opening, closing, dilation, erosion

Installation

pip install scikit-image

File Compression

Overview

Compress processed images using Zstandard compression for efficient storage.

Function: Compress with Zstandard

Suffix: _compressed (file becomes .zst)

Parameters:

  • remove_source (bool, default: False): Delete original after compression
  • compression_level (int, default: 3, range: 1-22):
    • 1-3: Fast compression, larger files
    • 4-10: Balanced
    • 11-19: Better compression, slower
    • 20-22: Maximum compression (ultra mode)

Use case: Save disk space for large datasets.

Installation

Zstandard must be installed system-wide:

Linux

Pre-installed on most distributions

macOS

brew install zstd

Windows

pip install zstandard

Usage

Run compression as a post-processing step after other processing functions.

Tips

  1. Compression Level: Level 3 offers good balance of speed and compression
  2. Remove Source: Enable only after verifying compressed files are valid
  3. Batch Processing: Compress entire folders efficiently
  4. Decompression: Use pzstd -d filename.zst to decompress

Colocalization Analysis

Overview

Analyze colocalization between labeled regions across multiple image channels.

Use Cases

  • Quantify overlap between different cellular markers
  • Analyze protein-protein colocalization
  • Study spatial relationships in tissue sections
  • Count nested structures (e.g., nuclei within cells)

Typical Workflow

  1. Multi-channel imaging: Acquire images with multiple fluorescent markers
  2. Segmentation: Label regions in each channel (cells, nuclei, organelles)
  3. Colocalization analysis: Quantify overlaps between channels

Expected Data Structure

Multi-channel label images where:

  • Channel 0: Reference channel (larger structures, e.g., cells)
  • Channel 1: Secondary channel (smaller structures, e.g., nuclei)
  • Channel 2: Optional tertiary channel (sub-nuclear structures)

Outputs

Statistical tables containing:

  • ROI counts per channel
  • Colocalization counts
  • Optional size measurements
  • Median/total sizes of colocalizing regions

Key Features

  • Boolean Masking: Determines overlap using binary masks
  • Unique Label Counting: Counts distinct objects in overlapping regions
  • Size Analysis: Optional measurement of ROI sizes
  • Multi-level Analysis: Supports 2 or 3 channels

Technical Notes

  • Reference channel typically contains larger structures
  • Other channels contain smaller, potentially nested structures
  • Each unique label ID represents one instance
  • Zero (background) is always excluded from analysis

See Also