Scientific Image Analysis
A workflow for segmenting, counting, measuring and analysing microscopy images, disease images, particle images and tissue sections.
Steps
Organise and back up images
File management / GIMPStandardise image names, formats, resolution and experimental groups, and keep a backup of the original images.
Never overwrite original images; save all processed results to a new folder.
Preprocess
ImageJ / FijiCrop, calibrate scale, subtract background, split channels and check brightness and contrast.
Use the same preprocessing parameters for images from different groups.
Segment and identify
ilastik / QuPath / ImageJChoose threshold segmentation, machine learning segmentation or object detection depending on the image type.
Spot-check segmentation results manually; don't simply trust the automatic output.
Measure and export
ImageJ / QuPathExport metrics such as area, count, length, grey value, fluorescence intensity or lesion proportion.
Keep image IDs and experimental groups in the exported table for statistical analysis.
Statistics and plotting
Python / RRun statistical tests on image metrics and visualise them to produce results figures for the paper.
Mind sample independence: multiple objects in one image are not necessarily independent replicates.
Articles & Guides
- Measuring Brinell indentation diameters with OpenCV: method, error and validation
Image calibration, segmentation, fitting and the HBW calculation for measuring Brinell indentations with Python and OpenCV, and why a 1% error in the diameter becomes roughly a 2% error in hardness.
- Error and validation in research image measurement: from calibration to agreement analysis
A number produced by an image program is not more reliable just because a program calculated it. This article walks through sources of error, calibration, repeatability and agreement with a reference method, and what a report should state.
- Scientific image measurement with ImageJ/Fiji: from calibration to exported results
Using a simulated image of known size, work through calibration, ROIs, thresholding, area and length measurement and export, then deliberately get settings wrong to see how far results move.