- Guide
- Imaging
- Tune an image analysis
Tune Cell Segmentation and Detection
An image analysis opens on one image with its pipeline in the sidebar. Tune the steps on that one image until the detection matches what you meant, then run the same settings across the set.
The sidebar keeps the same layout throughout, with Analysis run at the top for the runs you have made and Pipeline under it for the steps and the controls that run them.
Opening a Step
Open any step in the sidebar to see one line on how it works, then its settings. How it works under that line opens the full account, and the info button beside a setting's name explains that setting.
Changing one re-runs the pipeline live on the image in front of you. The settings are named for what they do, such as whether to measure bright or dark areas, an automatic or fixed threshold, a minimum object size, or whether to clean up background.
An open step lists every setting it has, the everyday ones first and the finer ones after them, such as how an automatic threshold is chosen, how hard to split touching objects, how far a blur or a cleanup reaches, and what a gate keeps objects on, shape included.
Some steps include a setting that no other step has.
| A step that | Its own setting |
|---|---|
| Cleans up object shapes | Remove objects touching the edge, which drops the objects the image edge cut off, so their clipped size and shape stay out of the results |
| Measures covered area | Threshold strategy, with every choice detection by brightness offers apart from Adaptive, because coverage records one cutoff for each image |
| Matches objects to the regions holding them | Match by. Center inside the region matches each object to the region its center falls in, and Most of the object inside matches it to the region holding the largest share of it, once that share reaches Smallest share inside |
| Scores overlap between two markers | Overlap thresholds. Automatic (Costes) picks the cutoffs for Manders' coefficients from each image, and Fixed cutoffs uses the Marker A cutoff and Marker B cutoff you set, so the same cutoffs hold across a set |
Remove step at the foot of a step asks you to confirm, then takes it out. A step that read it reads what it read instead, so removing Split touching objects leaves the count working on the objects detection found.
A step left with nothing to read comes out too, such as a range that keeps objects by a measurement you're removing. The confirmation names every step that comes out and every step that will read something else.
Removing the last step, or the step every other step works on, leaves the pipeline with no steps. The sidebar then reads Add your first step, and Add step starts it again. Earlier runs stay in the run history.
To change how objects are found and keep the steps below, use Replace step.
Replacing a Detection Step
Replace step sits at the foot of every step that finds objects in an image, and lists the other steps that do. The new step takes the old one's place, so every step below keeps working on the objects it finds.
It asks you to confirm when the swap would drop a setting you changed. On a plan that includes no compute usage, the steps that detect with an AI model are grayed out, with the reason on each.
Add step under the last step lists every step, with a search box to narrow them, and adds the one you pick at the end, open so you can set it. A step that needs an earlier step first is grayed out with a short note saying what's missing. Image analysis pipeline steps lists every step, what each one reads, and the settings it opens with.
Choosing What a Step Reads
When a step could read more than one thing, such as the whole image or the objects found, the image before or after a cleanup, or one of the image's channels, the choice sits at the top of its settings.
The list is in groups, so you can tell part of the image from the result of an earlier step.
- The image holds the channels by the names the microscope recorded, such as DAPI and GFP. A channel the file left unnamed reads as its position.
- All channels merged, last in that group, reads every channel drawn over the others in its display color. Point a step at a single channel when you mean one marker's pixels.
- A cleaned image from a step above holds each cleanup step above this one, by its name.
On a z-stack each entry says it measures the plane the viewer is on, so set Depth to the plane you mean to measure before you run the set.
Each step's line in the sidebar names the image it reads.
Detecting with an AI Model
Find colonies on a plate and Find fluorescent cells or nuclei detect with an AI model. Add either one from Add step, where both show an AI tag, or swap an existing detection step for one with Replace step. Each has two settings of its own, both starting at Standard.
| Setting | What it changes |
|---|---|
| Detection sensitivity | How much of the image comes back as objects. More objects picks up dim nuclei that Standard leaves out, and some background with them. Fewer objects keeps only the confident ones |
| Overlap between objects | What happens where two objects touch. Merge less keeps them apart, which suits a dense field of nuclei, and Merge more counts them as one, which suits colonies that have grown into each other |
Find objects by brightness cutoff has its own cutoff instead.
An AI step reads the image, or one of its channels, as it was acquired, so a cleanup step above it changes nothing about what it detects.
Changing either one and pressing Run detects the image again, which counts toward your team's monthly usage. An image already detected at the settings you are on is reused, so running it again adds nothing to that. Where the question is about size, confidence, or brightness, Filter under the image works on the objects a run already found and uses nothing.
Rearranging the Pipeline
To put a step somewhere else, click the grip at the start of the step it goes beside. Add step above and Add step below list the same steps, with the ones that can't go at that place grayed out and a short note on each.
A cleanup added above detection becomes the image detection reads. A step that measures brightness keeps reading the image as it was acquired, and you can point it at a cleaned image from its settings.
Move up and Move down in the same menu move a step one place, and dragging the grip moves it further. Two cleanups trade places along with what each one reads, and an object cleanup moved above a measure becomes the objects that measure reads.
A step always stays below the steps it reads, so the menu offers only the moves a step can make, and a cleanup that detection reads has no Move down.
Settings Left on Auto
A setting left on Auto is worked out again for every image, so its step shows the value it came to on the image in front of you, such as "Auto (Otsu), resolved to 118 on this image". A step that cleans up background shows "Background offset resolved to 30 on this image". Move to another image in the set to read its value there.
Expect threshold between, among that step's settings, records where you expect the value to fall across a set, such as 90 to 160. An image that falls outside the range is named in Checks with the value it reached, and the run keeps that value, so the range points you at the image worth opening and leaves the measurements as they were made. A background cleanup step has the same setting for its offset.
Leaving Out a Data Bar or Border
An SEM writes a data bar along the bottom of each image, and an automatic threshold reads that bright text with the specimen. Trim the image edges, first under Preprocess in Add step, leaves a band of pixels along any edge out of every step, and results are still drawn in place on the full image. A template includes the trims, so Run on all N images trims every image the same way.
Crop in the image viewer changes only how one image is shown, so reach for the step when the band should stay out of the numbers.
The step's settings say what the image on screen keeps, such as Keeps 1536 × 1024 of this 1536 × 1094 image. An image of a different size from the one the trims were set on runs with the same trims and gets a warning naming both sizes, which counts as a check on its card in the run review.
Try a Range
Try a range, beside a setting in the sidebar and beside the cutoff under the image, opens a panel under the image and tries values across that setting's range, drawing what each one measured.
Under the curve, a row of small pictures shows your image with each value's overlay on it, so you can see what each value does to the image. At the low end almost everything is marked, and at the high end almost nothing. The stretch where the number stops moving is colored, and a value in the middle of it still holds next week, when the exposure has drifted.
On a setting left on Auto, the row runs that value too and marks its picture Auto, with a pin on the curve above it, so you can see what the pipeline is doing right now among the values tried.
| Button | What pressing it does |
|---|---|
| Keep Auto, flag images outside 116–209 | Records the stretch as the expected range, and leaves Auto choosing for each image |
| Switch to Auto, flag images outside 116–209 | Does that, and releases a value you had fixed |
| Keep Threshold at 93 | Leaves a value you had fixed as it is |
| Set Threshold to 116 | Fixes the setting at one value |
| Set the cutoff to 40 | Moves a filter |
Point at a picture and the image above shows that value's overlay, so you can run along the row and watch the marks move on your own cells. Click one to choose it, or move the pick with the arrow keys.
Nothing changes until you press one of those buttons, so you can read the curve and close it.
Trying a range on a cutoff counts the objects the run already found, so it takes no time and uses nothing.
Trying a range on a setting that changes what is detected runs the pipeline again for every value, on the image in front of you. On a pipeline that detects with an AI model each of those runs would detect again and use that much usage, so Try a range stays off beside those settings, and hovering it points you at the cutoff under the image, which answers the same question for nothing.
The key under the image says what the overlay's colors mean for the step you have open, and on a measure step you hover a region to read its values.
Filtering What Counts
Under the image, Filter in the step's card trims what counts by Size and Brightness. A detection made by an AI model offers Confidence too, since the model scores each object it finds, and a run that sorted objects into two or more color groups offers Color. Filtering happens on results you already have, so narrowing a size range uses nothing and takes effect immediately.
Where the run measured it, Size offers px² or µm² and Brightness offers Raw or Above background. The second reading of each compares across experiments. µm² describes the same object at any magnification, and above background measures each object against the background of its own image, so a cutoff still holds when the exposure changes.
Measuring in µm² needs the image's scale set, and on a set where only some images have one, the images without it report a check saying so. Picking a unit clears that range, because the same number means something else in the other one.
On an image set it moves every image's saved count with it, counted again from the objects the run stored, so no run joins the history and nothing is spent, even where the pipeline detects with an AI model.
On a set, a range covers every image until you say otherwise.
Two buttons move a range between the set and one image.
- Filter this image separately narrows it to the image in front of you, and the panel then reads Filtering this image only.
- Use the shared range puts it back on the set.
Reach for a separate range where one slide was stained or exposed differently, and keep the shared one everywhere else, so the set is measured one way.
Some results cannot be counted again that way, such as a pipeline that counts objects inside a container, one that measures the whole image, or an older run. The card with Filter says which, in a line opening "The saved results still use the earlier filters", and Run on all N images beside it brings the table up to date.
Three controls put filters back.
- Show all in a panel clears that one filter.
- Reset to suggested, at the end of the row of filters, sets Size to the cutoff that hides specks far smaller than your main group of objects, when the image has them, and clears the other filters.
- Undo in the message that follows puts your earlier filters back.
Fixing What Looks Wrong
Fix a problem, beside Filter in the step's card under the image, lists what commonly goes wrong with a count and takes you to the setting that answers it. It appears on pipelines that count objects with the detection switches, such as Background cleanup and Split touching objects.
| Problem | What picking it does |
|---|---|
| Touching objects counted as one | Offers Split touching objects. With the split already on, it opens Try a range on Split depth |
| Specks or debris counted | Opens Filter on Size, where Trim to main group sits |
| Background counted as objects | Opens Try a range on the threshold |
| Faint objects missed | Opens Try a range on the threshold |
| Objects missed in dim corners | Offers Background cleanup |
| Holes, or one object in pieces | Offers Fill holes |
A problem that offers a switch tries it on the image in front of you first. Dashed outlines on the image show what it would count, and the line under the problems gives the count before and after, such as On this image: 8 → 12 objects.
Nothing in the pipeline changes until you press the button naming the switch, such as Turn on Split touching objects. That adds the step under detection, the same as turning the switch on in the sidebar. Cancel leaves the pipeline as it was.
On a pipeline that detects with an AI model, two problems point at the model's own settings.
- Touching objects counted as one offers Overlap between objects at Merge less.
- Faint objects missed offers Detection sensitivity at More objects.
Trying either would detect again and count toward your team's monthly usage, so these and Fill holes show no preview, and the count reflects the change after your next Run. The problems that point at the threshold or at background cleanup are left off the list there, since the model detects without either.
Ground Truth
Ground truth measures the pipeline against your own eyes. Count the objects on an image yourself with the counter in the viewer, then Use marks as ground truth on the annotation toolbar keeps those marks against the image.
Every run over it then reports how many of them it found and how many it found that you did not mark, beside the run's own number, under the run in the sidebar's Analysis run, and again in Accept run. Marks belong to the image itself, so a second analysis over the same image is measured against the same twenty cells, and they stay after the analysis that drew them is gone.
A run that disagrees badly with your marks is named in Checks.
Best match to marks appears beside Try a range on an image you have marked. It picks the value that agrees with your marks most closely, where Try a range looks for the stretch where a number stops moving. Reach for it on an image you have marked, and for Try a range everywhere else.
Checks
Each step reports its own Checks, which flag a step that found nothing or suspiciously many things, so you can catch the mistake before you run it across 200 images.
Checks also lists any image in the set whose automatic setting fell outside the range its step expects, naming the image and the value it reached, and that image's card in the run review shows a check of its own. A check is a warning or an error, so open any result with a count beside it. The panel appears only when a step reports one, and a clean run shows none. The values a step worked out for itself, such as a background offset, are in that step's settings.
A control image that missed its range is listed here too, at the severity that control was set to, which is flag as on its row in Controls.
Running the Set
The steps end in a Preview card. Run tries your settings on the image in front of you, and Run on all N images below it runs them across the whole set.
Trying Settings on One Image
- On a brightness-cutoff pipeline the image follows your settings as you type. A detector backed by an AI model waits for Run.
- Once these exact settings have run on the image, Run switches off until you change a setting. On an image set, the line under it sends you to the set run instead, since that is what saves numbers at new settings.
- A run over one image before the set has ever run is a trial, however many images you try. Each trial replaces the one before, stays off Analysis run, and your first Run on all N images is run 1. Close the image analysis after a trial and its numbers are kept.
Changing a detection setting changes only what's on screen. The results table keeps the numbers from the last run over the set until you run the set again, so you can try a dozen cutoffs on one image and leave the rest of your results alone. Moving a filter is the exception, because it counts again the objects the run already found.
Running Every Image
When the preview looks right, Run on all N images repeats the same pipeline across the set. For an AI detector, the card under the button says how many images still need detecting and how much of the month's usage is left, so a team with no usage left finds out before the run starts.
A pipeline with an AI detector keeps running after you close the tab, and a finished run appears in your inbox under Image Analysis Runs. Keep the tab open for every other pipeline. Keep it open too when an AI detector runs on any of these.
- a set of more than 50 images or slices
- a set of images inside a multi-image file such as a LIF
- a detection step that reads all channels together
Notify me when done raises a browser notification when a run finishes.
Once the set has run, Run on all N images switches off until you change a setting or add an image. Removing an image that has results asks you to confirm, then deletes that image's results from the results table, and every other image keeps its own.
Once a pipeline finds what you meant, Image Analysis Templates keeps those steps and settings under a name, so next week's plate is measured the same way.
Running on a Phone
The run sits at the end of the row of images under the photo. Run works through the image in front of you and shows how long it has taken, and once that image has a result the button reads Run on all N.
- For an AI detector, Run on all N first opens a card with the same usage line and meter as the sidebar.
- In an image analysis started from Plate Count, the photo starts detecting as soon as it finishes uploading.
- If a run fails, or the pipeline can't run on the image, the reason shows above the row.
Z-Stacks
A z-stack uses as much as its number of slices, so Conspecta checks which slices are in focus and offers to run only those. Depth, under the image, shows what it found, such as Slices 5–17 in focus · 13 of 40 will run, and you can accept it, widen it, use every slice, or reset to the suggestion.
Depth says one of five other things when the stack is not a single clean band, and the run follows whatever it says.
| What Depth says | What it found, and what runs |
|---|---|
| 2 regions in focus | Two stretches in focus with the blurriest slices of the stack between them, which are skipped. Both regions run unless you narrow it, and Depth offers Both regions beside a button per region, so a stack holding two specimens at different depths can run one of them on its own |
| Every slice looks usable | A stack thin enough that nothing is out of focus, so every slice runs |
| No clear focus | A stack that reads as uniformly dark, so there is no band to suggest and every slice runs |
| Focus check failed | The check could not score this stack. Every slice runs, and reopening the image analysis tries it again |
| your range | A band you set by hand, as in Slices 6–20 · your range · 15 of 40 will run. The suggestion stays one click away |
Across a whole set, Choose which slices lets you decide once whether to detect focus per stack or apply one range everywhere. The focus check itself uses nothing.
Each stack in that dialog box has a tag for what it came to, one of auto, nothing to trim, shallower, not checked, or a count such as 2 regions, so you can read what every stack came to before running any of them.
Once any image in the set is a stack, the run counts slices, so Run on all N images reads Run on 13 of 40 slices with the trim already in the number.
Watching a Run
A finished run opens onto its results. Running the set counts the images analyzed against the total with the time so far, and a finished image sits below it with its detections drawn on top, so you can see whether the detection is landing on your cells.
Analyzed so far is the row of everything the run has finished. Click a tile to hold that image and study it while the run continues behind you, and Jump to latest goes back to following the run.
Stop ends the run and keeps every image already analyzed, with its numbers. The results header then reads Stopped after 7 of 40 images, so the record shows how much of the set the numbers cover.
Editing an Image Analysis Someone Else Has Open
An image analysis saves as you work, and any number of people can work on the same one at once.
If you both changed the same part, the top of the image analysis shows Some fields were updated by another user. with each value under Keep mine and Use theirs, and nothing else you change saves until you choose and click Apply.
With no connection, a change to the title, description or status reads Saved on this device in the top bar and sends when you're back online. Steps, marks and the image list read Not saved with a Retry button until the connection returns, so keep the tab open.