1
Upload Models and Data
Bring images, masks, models and saved sessions in. One uploader takes the lot - files are sorted by extension, and models by their contents.
Upload Images, Masks and Models
File uploader
Takes images and masks (.tif, .tiff, .npy, .png, .jpg, .jpeg), model weights
(.pt, .pth), a Cellpose hyperparameters table (.csv) and a saved
session (.zip) - all at once if you like. Masks pair to images by filename;
Cellpose and DenseNet weights are told apart automatically. Unpaired or unrecognised
masks are ignored. 512 MB per upload, so split larger sets into batches.
Mask suffix
The text a mask filename adds to its image filename, so the two can be paired.
Default _masks, meaning sample1.tif pairs with
sample1_masks.tif. Change it to match your own naming.
Resize (512x512)
On by default. Shrinks images and masks to 512×512 on the way in, which keeps large
images responsive at some cost in detail. Cell measurements are still reported for
the original, pre-resize image.
Use demo data
Loads an example image set, so you can try the whole workflow without your own data.
Cellpose / DenseNet cards
Green MODEL LOADED with the filename once a model is in the session, orange
NOT UPLOADED otherwise. Both models are optional.
Uploaded images
Image table
Nothing is ticked by default here - tick only what you want to get rid of.
Remove selected images
Drops every ticked image, with its masks and labels, from the session. Greyed out
until something is ticked, and not reversible afterwards.
2
Annotate Images
The main workspace, in three columns: controls on the left, the image in the middle, zoom and pan on the right.
Image navigation & display
Left column, above the tabs.
Image n/N banner
Position and filename of the image you are looking at.
◀ / ▶ buttons
Unlabelled buttons either side of the slider. Step to the previous or next image.
Hidden when only one image is loaded.
Image index slider
Jumps straight to any image in the set. Also hidden for a single image.
Masks
Shows or hides the coloured mask overlay. Filled blue while on.
Normalize
Stretches pixel intensities across the full brightness range, making faint cells
easier to see. Display only - the stored image and every measurement are unchanged.
Image
Shows or hides the photograph underneath. Turn it off to inspect masks alone on
black.
Current Mode banner
At the top of both tabs. Names the mode the image is in, so you always know what a
click will do. In labelling mode it also names the class, e.g.
Assign class (spore).
Segmentation Controls Tab
Three pop-out panels. Both tabs stay loaded behind the scenes, so every keyboard shortcut works from either one.
Predict masks with Cellpose Popover
Automatic segmentation of a whole image at a time.
Select model
Which weights to segment with: Cyto3, Cyto2, Cyto,
Nuclei, or Fine-tuned Model - your own, uploaded or trained this
session. Defaults to the fine-tuned model whenever one is available.
Generate
Segments the image currently on screen.
Batch generate
Segments every uploaded image with the same model and settings.
Cellpose hyperparameters
Reveals the five accuracy settings below.
Mean cell diameter
Typical cell size in pixels. Leave at 0 to let Cellpose estimate it, or set
it yourself when the estimate looks off.
Cell probability threshold
−2 to 2. How sure Cellpose must be that a pixel belongs to a cell. Higher gives
fewer, tighter cells; lower gives more, larger ones.
Flow threshold
−2 to 2. How irregular a predicted shape may be before it is thrown out. Lower is
more permissive.
Minimum cell size
Discards masks smaller than this area - a quick way to drop debris.
Niter
0 to 2000 refinement iterations. Higher values favour long, stringy cells such as
filaments.
Predict masks with boxes Popover
SAM2, for segmenting chosen cells one at a time.
Draw box
Click and drag a red box around each cell you want segmented.
Generate masks from boxes
Runs SAM2 on every box drawn, turning each into a mask. The boxes clear once their
masks are in.
Manually edit masks Popover
Draw masks by hand, and repair ones that came out wrong.
Freehand
Click and hold to trace an outline; releasing closes it into a filled mask.
Ellipse
Drag a box around a colony and it fills with an ellipse inscribed in that box - a
circle if you drag a square. Committed the moment you release.
Split masks
Drag a line right through a mask to cut it in two. The line has to cross all the
way through - partly crossed masks are left alone. Both halves keep the original
class.
Join masks
Lasso a group of masks; those lying entirely inside the lasso and touching
each other merge into one.
Classification Controls Tab
Manage Labels Popover
Classes are shared across every image in the session.
Add a new class:
Type a name and press Enter. The class is created, given a colour, and made the
current one.
Delete a class:
Removes a class everywhere. Cells that carried it become unlabelled - the masks
themselves stay. No label is reserved and cannot be deleted.
Rename:
Picks which class to rename.
To:
The new name; press Enter to apply it across every image at once. Typing the name
of a class that already exists merges the two.
Map model predictions to cell classes
Only present once a DenseNet model is loaded. One Map model class n to
dropdown per model output, tying the model's numbered outputs to your class names.
Until at least one is mapped, the classify buttons stay disabled.
Classify cells with Densenet Popover
Classify
Labels every mask in the current image with the loaded model. Disabled - with a
tooltip saying why - when no model is loaded or nothing has been mapped yet.
Batch classify
The same across every uploaded image, with a progress bar. Batch runs are not
recorded for undo.
Class rows List
One row per class - a colour chip matching the overlay colour, the name, then two buttons. No label sits first, and is how you strip a label back off.
All
Assigns every mask in the current image to that class in one go.
Click
Makes that class active and switches the image into Assign class mode, so
clicking a cell labels it.
Editing tools
Below the tabs, and available from both.
Remove
Click a mask to delete it, or click a drawn box to delete the box. Boxes take
priority where the two overlap.
Clear All
Wipes every mask, label and box from the current image. Other images are untouched,
and Undo will bring it back if you press it straight away.
Undo
Reverts the last mask or box action. Single-level - only the most recent action can
be recovered.
The image itself Canvas
Click and drag behaviour
Exactly one mode is active at a time, named in the Current Mode banner and set by the
mode buttons above. Remove and Assign class respond to a single
click; the rest are drag actions.
Chart toolbar
In drag modes a small toolbar appears above the image with the one tool that mode
uses - box select, or lasso. Nothing to choose; it reflects the mode you are in.
Scroll wheel
Deliberately off, so the view does not jump while you work. Use the Zoom slider, or
your browser's own pinch-zoom.
Zoom & pan
Right column, beside the image.
Zoom
1× to 10× in steps of 0.5, magnifying around the current view centre. The image stays
the same size on screen - you simply see less of it. At 1× the whole image is shown.
Minimap
Thumbnail of the whole image with a red rectangle marking the part you are viewing.
Click anywhere on it to recentre the view there.
▲ ◀ ◎ ▶ ▼ pad
Nudges the view by most of a screenful, keeping a sliver of the old view to orient
by. ◎ recentres on the middle of the image. All disabled at 1×, where
everything is visible anyway.
3
Train Models
Fine-tune a model on what you have annotated. Numbered steps; the parameters in step 3 depend on which model you picked in step 1.
Step 1 - Choose which model to fine-tune
Model type
Pick one: Train a Cellpose model to identify cells for better segmentation,
or Train a DenseNet model to classify cells for better labelling. Everything
below re-renders to match.
Step 2 - Select images to train on
Image table
All images ticked by default; untick any to exclude. Cellpose uses the masks on the
ticked images, DenseNet uses their labelled cells.
Step 3 - Cellpose segmenter parameters
Base model
Weights to start from - cyto, cyto2, cyto3,
nuclei - or scratch to train from random initialisation.
Max epochs
1 to 1000 passes over the training set. More can improve results, at the cost of time.
Learning rate
Step size the optimizer takes. Lower trains more slowly but more stably.
Weight decay
Regularisation over the final epochs, discouraging the model from over-fitting to
your training images.
Min cells per image
Images with fewer annotated cells than this are dropped from training - Cellpose's
own default is 5. The step 4 summary reports how many this excludes.
Batch size
8, 16, 32 or 64 images per training step. Larger is faster but needs more memory.
Images per epoch
How many training images to sample each epoch - All, or 8 / 32 / 64 / 128 / 256.
Optimise hyperparameters
After training, searches for the inference settings - diameter, thresholds and so on -
that suit the new model best. Note the winning values down: you will want to re-enter
them on the Annotate Images page when you use the model.
Optimisation iterations
10 to 60 combinations to try, in steps of 5. Only shown when the box above is ticked.
More trials, better settings, longer wait.
Train / test split
10% to 90% in steps of 5, labelled with the share of images used for
training. The rest are held out as a test set and never trained on, so the step 5
plots measure performance on images the model has not seen.
Step 3 - DenseNet classifier parameters
Input size
64, 96 or 128 pixels - the size each cell crop is scaled to before the model sees it.
Larger keeps more detail and trains more slowly.
Batch size
As above, but counted in cell crops rather than whole images.
Max epochs
1 to 500 passes over the labelled cells.
Train / validation split
Works like the Cellpose split, but divides labelled cells. The held-out
share becomes the validation set behind the accuracy and confusion matrix plots.
Step 4 - Train the model
Only one model trains at a time; starting one greys out the other.
Training set summary
Sits just above the button. Cellpose reports total masks, images passing the
minimum-cells filter, and the train/test breakdown. DenseNet reports total labelled
cells, the train/validation breakdown, and a per-class count table.
Fine-tune Cellpose
Starts training in the background.
Fine tune Densenet121
The same, and additionally disabled until at least two classes have two or more
labelled cells each.
View Training Log
Appears while a job runs - live worker output, so you can follow progress epoch by
epoch. Cellpose shows a matching View Validation Log during its evaluation
pass afterwards.
Cancel Training
Stops the running job; nothing is kept from a cancelled run. Reads
Cancel Validation during the evaluation pass.
Show Error Details
Appears if a job fails, holding the full error from the worker.
Step 5 - Review model training
Appears once a model has finished.
How do I interpret these graphs?
Plain-language notes on each plot, one expander apiece - training vs test loss, IoU
comparison and predicted vs real counts for Cellpose; loss, accuracy / precision /
recall / F1, and the confusion matrix for DenseNet.
Tested hyperparameters
Shown when Optimise hyperparameters was ticked: every combination tried and
how it scored, so you can copy the best into the Annotate Images page.
A newly trained model goes straight to work. As soon as training
finishes it becomes the Fine-tuned Model option on the Annotate Images page,
and the Cellpose model / DenseNet model tick on the Downloads page.
4
Compare Phenotypes
Plot the shape and colour descriptors of your segmented cells, grouped by class. Nothing is calculated until you press the button in step 5.
Step 1 - Select images to compare
Image table
All ticked by default. Untick to plot a subset - one condition against another, say.
Steps 2 & 3 - Choose what to plot
Choose classes to compare
The classes present in the ticked images, plus Unlabelled for cells with no
class. Each becomes one group in the plots; remove one with the × on its tag.
Choose cell descriptors to compare
Pick as many as you like - area, perimeter, circularity, solidity, eccentricity and
the rest, colour measures included. One plot comes out per descriptor.
area is selected to begin with.
Attribute Information
An expander per descriptor explaining what it measures and how to read it, with a
diagram.
Step 4 - Select plot parameters
Plot type
Violin shows the full distribution per class; Bar shows the summary.
Overlay datapoints
Draws every individual cell over the plot, so the sample size and any outliers behind
the summary are visible.
Convert to distance
Turns pixel measurements into real units, and reveals the two controls below.
Pixel size
The physical size of one pixel, in whatever unit you name alongside.
Unit
Labels the axes - µm gives “Area (µm²)”. Defaults to µm.
Step 5 - Generate plots
Generate Plots
Builds the measurement table and draws one plot per selected descriptor. Press it
again after changing anything above to refresh them.
5
Downloads
Tick what you want across four columns, build the zip, then save it. Changing any tick clears a zip you already built, so you always get what is currently ticked.
Images and Masks
Images and masks
On by default. The image files and their mask files. Unticking it greys out the four
options below, which all build on it.
Colored mask overlays
On by default. Copies of each image with the cells drawn over them in their class
colours - the view you see in the app.
Per-image class counts
Prints how many cells of each class there are into the corner of each saved image.
Normalize images
Rescales pixel intensities to the full 0-255 range before saving.
Cell patch images
A small cropped image of every individual cell - useful as training data or for
figure panels.
Tables
Per-image cell counts
On by default. A CSV of how many cells of each class appear in each image.
Cell metrics
On by default. A CSV with every morphological and colour descriptor for every single
cell - the raw data behind the page 4 plots.
Trained Models
Cellpose model
Fine-tuned weights, the training dataset and the loss curves. Greyed out until a
model has been trained or uploaded.
DenseNet model
The same, with evaluation metrics in place of the loss curves.
Session Restore
Save Session
Adds mycol_saved_session.zip to the download. Drop that zip back into the
uploader on page 1 to restore your images, masks, labels, models and settings exactly
as you left them.
Building the download
Prepare Download
Packs everything ticked into a single zip. Large sets take a while - a spinner shows
while it works.
Download Files
Saves the prepared zip to your computer. Greyed out until Prepare Download
has been pressed.
⌨
Keyboard shortcuts
All on the Annotate Images page, and all working from either control tab.
← → Previous / next image
Space Toggle mask overlay
N Toggle normalization
I Toggle underlying image
B Draw box mode
G Generate masks from boxes
F Freehand mode
E Ellipse mode
S Split masks mode
J Join masks mode
D Remove mode
Ctrl+Z Undo last action
Shortcuts know when you are typing. Letter and Space shortcuts
stand down while the cursor is in a text box, so naming a class never triggers a mode
change and a space stays a space. The zoom
slider hands the arrow keys back as soon as you move away from it, so ← and
→ keep changing image.