Bioimage Quantification and Analysis Notes

Introduction

Bioimage quantification involves extracting meaningful information from biological images using image processing techniques.
It is important because it reduces selection bias, is faster, provides high throughput for greater statistical power, and gives a better descriptor of experimental results.

Image Acquisition

Optimizing experiments, choosing the best equipment and settings, and observing proper sampling are crucial for image acquisition.
Proper sampling relates to meeting the Nyquist criterion, which applies to both spatial and temporal resolution, and bit-depth.
For imaging, slight oversampling (2.3 to 3x) is advised.

Microscopy Techniques

Various microscopy techniques are used in cell biology, including Brightfield, Fluorescence (Widefield, Point scanning), specialized techniques, and Electron microscopy.

What is an Image?

An image is a matrix of pixels, with each pixel having parameters like XYZ position, size, and intensity.
Pixel intensity is proportional to signal strength.
Permutations of these basic parameters are used in image analysis.

Qualitative vs. Quantitative Image Analysis

Qualitative image analysis involves manual counting, minimal processing, small sample sizes, and potential selection bias, leading to low throughput and statistical power.
Quantitative image analysis involves automated processing, reduced selection bias, faster throughput, greater statistical power, and better descriptors of experimental results.
The brain is good at identification but not quantification due to perceptual illusions like the Ebbinghaus illusion and Hermann Grid.

Optimizing Experiments

Appropriate fixation, permeabilization, and labeling are essential as they affect morphology and protein localization.
Use the best equipment and settings possible; for example, the Numerical Aperture (NA) of the objective lens affects resolution and intensity.

Proper Sampling

The Nyquist criterion/Nyquist-Shannon sampling theorem dictates that to accurately represent spatial resolution in digital images, the sampling interval must be at least twice the highest spatial frequency.
Undersampling occurs when the detection frequency ( $f'$ ) is less than the signal frequency ( $f$ ).
Oversampling occurs when the detection frequency ( $f'$ ) is greater than the signal frequency ( $f$ ); aim for $f' = 2f$ or $f' = 4f$ although $f' < 8f$ still provides oversampling.
Resolution is determined by the objective NA and wavelength of light; adjust your pixel size as required.
Must have at least 2 pixels per resolvable structure (2.3 to 3x is better) to adhere to Nyquist sampling in digital images.
The Nyquist criterion also applies to time intervals; for example, a cell with a diameter of 10 um that moves at 5 um/minute can be undersampled.

Image Bit-Depth

Higher bit-depths provide more grey values, resulting in more detail in both structure and intensity.

Binning

Binning boosts camera frame rate, dynamic range, and signal-to-noise ratio by sacrificing resolution.
It combines data from adjacent pixels into a super pixel rather than reading out each individual pixel.

Bit-Depth Examples

3-bit: 8 grey levels, shorter exposure times, undersampling.
8-bit: 256 grey levels.
16-bit: 65536 grey levels, longer exposure times, oversampling.

Image Analysis Workflow

Image analysis extracts meaningful information using image processing techniques.
The workflow includes Pre-Processing, Segmentation, Feature Detection, and Data Presentation.

Pre-Processing

Involves reduction of acquisition artefacts that could influence results, such as shading and noise.
Improves image quality for better target identification through edge enhancement.
Examples include Smoothing, Sharpening, and Deconvolution.

Segmentation

Partitions the image into different segments based on image information such as intensity, color, pattern, or pixel position.
A binary process assigns 1 to the object of interest and 0 to the rest, resulting in a binary image that can be further refined by operations like dilation, erosion, or separation.

Segmentation Techniques

Threshold (based on intensity): Suitable for fluorescence images.
Variance (local change in intensity): Suitable for brightfield images.
Machine learning: Can be used universally but is computationally heavy.

Threshold-Based Segmentation

Traditional thresholding, such as Otsu thresholding, works well for many applications.
Successful segmentation depends on camera specifications, image quality, complexity, consistency, and image type.
Advantage: Fast.
Disadvantages: Requires high contrast and isn't very specific.

Machine Learning Image Segmentation

Training involves using a subset of images labeled by a human to extract features and train a classifier, which is then used to segment unlabelled images.
Machine learning makes it easy for non-experts to use ML with simple GUIs.
Examples include segmenting images with non-uniform background illumination, classifying cell compartments, and analyzing scratch assays and muscle sections.

Feature Detection

Involves measuring parameters for single objects, groups of objects, conditional objects, and their associations.
Examples include measuring protein production inside cells or measuring dividing cells.

Measured Parameters

Area: $\sum$ of pixels in segmented object; object vs. total; cell/tissue coverage.
Total intensity: $\sum$ of all pixel intensities in object; expression level/amount.
Average intensity: Total intensity/area; concentration of signal/dye/protein.
Major Axis (Length): Longest line drawn in object.
Minor Axis (Width): Longest line perpendicular to length/major axis.

Texture Analysis

Texture is analyzed alongside integrated and average intensity.

Data Presentation

Line chart for time series (e.g., scratch assay).
Histogram for distribution analysis (e.g., cell size distribution).
Scatter plot for relationship analysis (e.g., gene expression).

High-Throughput Screening (HTS)

Extends from single cell analytics (n=1) to multi cell analysis (n=10,000).
Screens large numbers of variables like drug libraries, concentrations, and interactions.
Robust analysis parameters include cell numbers, apoptosis, and neurite outgrowth.
Usually involves robotics.

HTS Case Study

Effect of drugs on mitochondria morphology using PerkinElmer Opera® HCS System and Columbus Analysis software.
High degree of phenotypic variability, requiring analysis of 10,000s of cells to detect subtle effects of treatments that human investigators might miss.

Multi-Wavelength Cell Scoring

Cell counting based on multiple wavelengths.
Objects (nuclei, cytoplasm) identified with Min/Max widths and thresholds.
Measurements include Scoring (+ve/-ve cells), Area (area/per wavelength), Total intensity, Average intensity (per cell or image).

Cell Cycle Analysis

Assign cell cycle stage based on the amount and concentration of DNA present, similar to flow cytometry.

Custom Modules

Analysis modules are good at doing their own task well, but custom modules are necessary when complex multiparametric analysis is required.

Ilastik

Ilastik is a free open-source software for image classification and segmentation, requiring no previous experience in image processing.
Optimize imaging (WGA selection, Optimize labelling and Microscopy parameters. Good even exposure, no reduction in contrast through the cell, select samples clearly visualized, no cut cells. Exclude noisy images/sections.

Ilastik Workflow

Includes raw confocal image, feature selection and training, self-probability learning, and binary output image.
Can be trained on a different set of images with varying quality/parameters.

Example

Post-processing steps in image analysis include filling holes and analyzing particles based on size.

Pirfenidone Study

Pirfenidone increases transverse tubule length in the infarcted rat myocardium, as quantified through tubular density and area measurements.

Summary

Optimise experiments : Good sampling
Understand methods : Use appropriate techniques
Quantify data , preferably automatically
Robust data = good science