Abstract:

This video is a SciPy 2021 tutorial on the fundamentals of bio-image analysis using Python. It covers a standard workflow, starting with basic image filtering concepts and building toward a complete segmentation and analysis pipeline for a 3D cellular dataset. The tutorial uses scikit-image for core algorithms like convolution, filtering (Gaussian, median), and edge detection (Sobel). It introduces napari as a tool for interactive, multi-dimensional visualization. A key part of the tutorial demonstrates a marker-based watershed segmentation to separate touching objects. Finally, it introduces dask and dask-image to address the problem of processing datasets that are too large to fit into memory, showing how to apply filters to out-of-core 4D time-series data.

A Python Tutorial on Bio-Image Analysis

• 00:05:29 Filtering Fundamentals: The tutorial begins by using 1D signals to build intuition for image filtering. It covers basic concepts like convolution for smoothing (mean, Gaussian filters) and edge detection (difference filters) using numpy and scipy. A key takeaway is the importance of handling edge effects and smoothing an image before attempting to find edges.
• 00:28:21 Filtering in 2D: The concepts from 1D are generalized to 2D images, which are treated as NumPy arrays. The tutorial demonstrates using scikit-image for common filters, highlighting that the median filter is often superior for denoising as it preserves sharp edges better than mean or Gaussian filters.
• 01:11:37 Interactive Segmentation with Napari: The focus shifts to segmentation—identifying and separating objects. The tutorial introduces napari, an interactive viewer, to visualize a 3D image of cells.
• 01:24:24 Thresholding and Morphological Cleanup: A basic segmentation is performed by applying a threshold to create a binary mask of the cells. This mask is then cleaned up using morphological operations from scikit-image to remove small objects and fill holes.
• 01:30:30 Marker-Based Watershed: To solve the common problem of touching objects being merged into one, the tutorial implements a marker-based watershed algorithm. It finds object centers (markers) by calculating a distance transform on the binary mask, smoothing it, and finding the local maxima. This is an effective method for separating clustered objects.
• 01:37:43 Measuring Object Properties: Once objects are segmented and labeled, skimage.measure.regionprops is used to extract quantitative features like area, intensity, and solidity. The results are loaded into a pandas DataFrame for analysis, which is used to identify an outlier (a dividing cell) based on its properties.
• 02:00:59 Handling Large Datasets with Dask: The final section addresses the problem of datasets too large to fit in RAM. It introduces dask for out-of-core, parallel computation on chunked arrays. When used with napari, Dask allows for the visualization of massive datasets by only loading the visible chunks on demand.
• 02:08:31 Applying Analysis with Dask: The tutorial demonstrates applying filters from dask-image and using dask.array.map_blocks to apply arbitrary scikit-image functions to each chunk of a large 4D dataset. It also shows how to visualize cell tracking data using napari's Tracks layer.