Skip to content

Latest commit

 

History

History
135 lines (99 loc) · 5.27 KB

puncta_segmentation.md

File metadata and controls

135 lines (99 loc) · 5.27 KB

Puncta segmentation

The puncta segmentation procedure consists of the following steps:

  1. Laplacian of Gaussian (LoG) filter to identify potential puncta centers

  2. Filtering of the potential puncta centers by comparing their intensities to the background signal of the cell

  3. Thresholding of the image either in the original intensity space, or the LoG filtered space

  4. Watershed segmentation of the thresholding result using the filtered puncta centers as seeds

  5. Optionally, excluding puncta that are too big or that extend beyond the cell mask.

Step 1: Detection of puncta center candidates

First, the candidate puncta centers are detected using the scikit-image implementation of the Laplacian of Gaussian (LoG) filter, which is a common blob detector.

This implementation allows the user to specify the minimum and maximum size of the target puncta, number of size scales applied, option to remove overlapping centers, and a threshold to remove low intensity centers (specified in the LoG filter scale).

All these parameters can be specified by the user and adjusted using the setup_puncta_analysis.ipynb notebook with immediate feedback on the result.

Step 2: Puncta center filtering

If the cell or nuclei segmentation masks are available, there is an option to additionally filter out puncta centers based on their intensities relative to the background intensity of the cells.

The background intensity of each cell is calculated as a specified percentile (default is 50: the median) of the image intensity in the cell (of the channel in which the puncta are detected).

If the “global_background” parameter is set to False, individual background values are used for each cell. Otherwise, the global background is calculated as a specified percentile (default is 95) of the background intensities of all cells in the image.

Image intensity at each puncta center is compared to the background intensity, and the center is removed if its intensity is lower than the background multiplied by the “threshold_background” (specified by the user, default is 3).

Step 3: Segmentation of puncta from the background

Next, to segment the puncta from the background, the puncta channel is thresholded in either original intensity space or LoG space.

The method of thresholding is specified by the “segmentation_mode” and “threshold_segmentation” parameters. The “segmentation_mode” takes value 0, 1, or 2 and determines the way the “threshold_segmentation” is applied.

For mode 0, the image is thresholded in the LoG space, i.e., the result of the LoG filter is used for thresholding; the segmentation threshold is applied as the absolute value, i.e., all pixels with LoG intensity higher than the threshold are assigned to the foreground / puncta.

For mode 1, the image is thresholded in the LoG space, as in mode 0, but a relative threshold is applied; for this, the background LoG signal is calculated in each cell similarly to step 2, and the LoG image is thresholded at the value of the “threshold_segmentation” multiplied by the background value.

For mode 2, the thresholding is done for the raw fluorescent channel, and the threshold is calculated relative to the cell background signal calculated in step 2.

The background for steps 1 and 2 is calculated globally, and the same value is used for all cells (note: this is a different background than the one used in step 2 for filtering of puncta centers).

It should be noted that since modes 1 and 2 require background calculation, they should only be used if the cell / nuclei segmentation is available.

The value of the “threshold_segmentation” parameter should be chosen depending on the value of the “segmentation_mode” parameter, since it has different meanings for different modes.

Thus, for mode 0, the value should be very small and close to 0, either positive or negative; we recommend starting with values 0.001 – 0.003 and decrease or increase the value depending on whether more/bigger or fewer/smaller puncta should be detected.

For modes 1, and 2 the “threshold segmentation” value specifies how much higher than the background the puncta intensity should be. For mode 2, a good value range to start with is 2 – 3.

Since the LoG filer increases the image contrast between the puncta and the background, the threshold value for mode 1 should be much higher; we recommend choosing a value from the range 20 – 100.

Step 4: Splitting puncta clusters

After the puncta are thresholded from the background, we split touching puncta using distance transform watershed, seeded with the puncta centers from step 2.

This allows splitting puncta that were identified as one region after the thresholding in step 3.

Step 5: Filtering of segmented puncta

Finally, we provide an option to exclude puncta that are too big or that extend beyond the cell mask.

If the option “remove_out_of_roi” is set to True, the puncta and puncta parts that extend beyond the cell masks are removed.

If a value for the maximum radius (maxrad_um) is provided, all puncta are removed that have a volume greater than the volume of the equivalent sphere calculated from the maximum radius.