Remove Objects

Both engines live in a layer's Post-Processing section and act on the events the layer has already detected; the reference lists no input image for either. Remove objects v2 holds a list of criteria. Each criterion is one measurement with a lower and an upper limit, and its own enable box. An event inside the range survives, and one below the lower limit or above the upper limit is removed.

Remove objects touching border has a single Enabled switch. It removes events that touch the border of the sample or region of interestLoading..., which represent incomplete structures. In its reference example, the structures of an IHC colon sample that touch the border drop out and the rest remain.

The Basic Operations Module has smaller relatives, described on the Removal operationsLoading... page. Remove labels (larger), (smaller), (stronger) and (weaker) each apply one threshold, and Remove objects (with seeds) and (without seeds) keep or drop objects by whether they overlap a second image. The NucleiLoading... engine also has removal built in, by percentage and by fixed limits.

Removing events can leave gaps in the label list. The Engines Reference recommends the Re-Label (missing)Loading... operation to correct it, and Existing Coded offers a Check and fix labels option for events imported from another layer.

Filtering by property, in general. Image processing has two broad ways to remove unwanted objects. Morphological openingLoading... removes anything too small to hold the structuring elementLoading..., so it tests size and shape through a probe. Filtering by measured property tests each object by its numbers, any number you can measure. Remove objects v2 is the second kind.

Shape tells debris from cells. Area alone can't tell a round nucleus from a streak of stain of the same area, but a shape measurement can. Standard Measurements defines Compactness as how closely an event resembles a disk: near 1 for a disk, near 0 otherwise. A common textbook formula for that idea is 4π × Area / Perimeter², on which elongated debris scores low. The Engines Reference shows Remove objects v2's measurement list only as a figure, so check in the software which shape measurements it offers.

Remove before you count. In statistics generally, a handful of gross outliers can shift a mean and inflate a variance. False detections do the same to population statistics, and they can land inside a gate and pose as a rare population. Removing obvious non-cells in Post-Processing keeps them out of every number downstream, including the classes assigned by Assign classes to objectsLoading....

Don't remove the biology. Removing too much does as much harm as removing too little. A lower area limit set for tumor nuclei can remove small lymphocytes, and an intensity limit can remove real cells in a dimly stained region. Look at what each criterion removes on a few fields before you apply it to a cohort.

Turn one criterion on at a time and look at what disappears. A limit that removes only debris on one sample may cut into real cells on the next, so check a few samples from across the cohort.

Source: StrataQuest 8.1 Engines Reference, ch. 37 (pp. 504–505), ch. 38 (pp. 506–507), ch. 6.9 (pp. 284–300), ch. 6.6 (p. 198), ch. 21 (p. 431), ch. 41 (p. 538).

Suppose a nuclei detection on an autofluorescent tissue section has picked up debris along with the nuclei.

1. Add Remove objects v2 in the layer's Post-Processing section.
2. On the Area criterion, set the lower limit just below your smallest real nuclei and the upper limit above your largest, so specks and merged clumps fall outside.
3. On Mean Intensity, raise the lower limit until faint autofluorescent blobs drop out, and stop before dim real nuclei do.
4. If the measurement list offers a shape measurement, add it as a criterion to drop elongated streaks.
5. Add Remove objects touching border so cells cut by the region's edge aren't counted as whole cells.

Check the overlay after each step. The events that remain are the ones Standard MeasurementsLoading..., the diagrams and the phenotypes will use.