← All projects

Image analysis ยท Biomedical

Cell Coverage Estimation

How much of a fluorescence microscopy image is covered by cells? A simple Otsu-threshold approach, measured against manually corrected reference masks to see how far a simple method gets.

Role
Solo
Data
BBBC008v1: human HT29 colon-cancer cells
Type
Image analysis
Stack
Python, Otsu thresholding
The short version~20 sec read
Problem

Estimate cell coverage automatically, and measure how accurate a simple segmentation method really is.

What I did

Applied Otsu's threshold to 12 actin-channel images and compared predicted coverage with expert reference masks.

Result

A mean absolute error of 19.7 percentage points. In the worst image it predicted 1.3% coverage against a true 41.8%.

01Method

  1. Load 12 grayscale fluorescence images of human HT29 cells.
  2. Use channel 3, which highlights actin in the cell cytoplasm.
  3. Apply Otsu's method to separate brighter pixels from the background.
  4. Calculate coverage as the percentage of pixels classified as cells.
  5. Compare with the corresponding reference mask.
  6. Compute the absolute error per image and the mean absolute error across all 12.
  7. Visualise the image with the largest error.

Coverage (%) = cell pixels ÷ total pixels × 100

02Results

12images evaluated
19.7ppmean absolute error
40.5pplargest single error
Three panels: the original microscopy image, the Otsu predicted mask showing only a few bright spots, and the reference mask showing much larger cell regions.
Fig. 1: The worst case. Original image, Otsu prediction (1.28% coverage) and reference mask (41.77%).

03What went wrong

In the worst image, Otsu's method picks up only isolated bright regions and misses most of the dimmer cell bodies. It thresholds on brightness alone, so cells with low fluorescence fall below the cut-off and coverage is badly underestimated.

04Limitations

  • Otsu relies on pixel brightness and struggles with low-intensity cells and dim boundaries.
  • Coverage error compares total area, not whether individual pixels are classified correctly.
  • Results describe the photographed fields of view only. They can't establish growth rates or coverage across a whole culture flask.

05Data & credits

Images and reference masks are from BBBC008v1 (Broad Bioimage Benchmark Collection), credited to David Root and Anne Carpenter, and licensed CC BY-NC-SA 3.0. References: Carpenter et al. (2006), Genome Biology; Ljosa et al. (2012), Nature Methods.

Next project NHS Antibiotic Prescribing Outlier Detection