Tech note: AI cell counting — a usable model from a single original image
Five crops from one image, each rotated ten times, gave 50 training samples; 30 images with 500 annotations trained an object-detection model. Classic image processing counted 1,637 cells in under 6 seconds.
Cell counting usually suffers from few samples, clumped cells and slow, tiring manual counts. Shin Shou tested two approaches for a research institute: classic image processing and deep learning.
Approach 1: classic image processing
- 200× image: 1,637 cells, average area 312 px, 45.7% area coverage, computed in 5.86 s
- 100× image: 9,634 cells, average area 57 px, 44.5% area coverage, computed in 20.78 s
Area coverage is reported too, useful for tracking culture confluence.
Approach 2: deep learning
- Generate samples: with only one original image, five random 500 × 400 crops were each rotated ten times, giving 50 samples
- Annotate: every cell labelled by hand; the object-detection model used 30 images with about 500 annotations
- Train: watch the training curves and tune parameters
- Validate: run images not used in training and adjust the training set based on the results
Once trained, the model can pre-annotate new images for human review, so labelling gets faster and faster.
Which to choose
When cells are well separated with clear contrast, classic processing is fast and stable; when they overlap or vary in shape, AI is more flexible. They also combine well: screen quickly with image processing, then let AI handle the difficult regions.
