Cell Well Segmentation

Requirements

• Windows executable release, or Python environment for source execution.
• Microscopy images in TIFF, OME-TIFF, or supported raster formats.
• For validation, a ground-truth GeoJSON annotation in the same coordinate system as the image.

Recommended use

Start with a small ROI in Parameter Exploration, tune the segmentation visually and with DICE/IoU if ground truth is available, then run the full image or batch set.

Main features

• Single-image and bulk-image processing.
• ROI-based parameter exploration before full processing.
• Nuclei-based seed detection and marker-controlled watershed segmentation.
• QuPath-compatible GeoJSON export.
• Skip completed or resume missing outputs for long batch runs.
• Optional DICE/IoU validation using ground-truth GeoJSON.

Channel convention

• 0: nuclear channel, for example DAPI.
• 1: red marker channel.
• 2: green marker channel.
• 3: cytoplasmic channel used for CellCyto segmentation.

Channel mapping can be changed in the GUI under Custom parameters.

Outputs

For each processed image, the app creates a dedicated output folder with:

• RGB.tif and CellCyto.tif composite images.
• instances.tif instance segmentation mask.
• cell_features.csv cell-level morphology and intensity features.
• manders_features.csv Manders colocalization metrics.
• cell_features_with_manders.csv merged feature table.
• manders_summary.json threshold summary.
• qupath_final.geojson QuPath-compatible annotations.
• preview.png quality-control visualization.

Cell-level features

The CSV feature table includes per-cell measurements such as:

• Label ID, area, perimeter and centroid coordinates.
• Red, green and blue mean, max, median, standard deviation and percentile values.
• Total channel intensity per cell.
• Coefficient of variation for red, green and blue channels.
• red_positive, green_positive and double_positive biological positivity flags.

Manders colocalization

Manders coefficients are computed per segmented cell. The red-in-green and green-in-red values are asymmetric and should be interpreted independently: one reports the fraction of red signal overlapping green-positive pixels, while the other reports the fraction of green signal overlapping red-positive pixels.

Biological red/green thresholds are used to classify cells as positive or double-positive and are saved for reproducibility; they do not directly change the segmentation mask.

Validation and parameter tuning

When a ground-truth GeoJSON is available, Cell Well Segmentation can compute pixel-level DICE, IoU, precision and recall. During Parameter Exploration, the full-image GeoJSON can be selected and the application evaluates only the selected ROI, making it easier to tune parameters before processing the full image.

Existing output handling

• Reprocess from zero: recomputes the selected images.
• Skip completed: avoids processing images with complete output folders.
• Resume missing outputs: regenerates missing outputs when intermediate files are available.

Notes and limitations

• Designed for fluorescence cell images where individual cells are distinguishable.
• Parameters may need adjustment for different magnifications, staining patterns or cell densities.
• Not intended as a replacement for manual expert review when validation or publication-quality analysis is required.