Binary + composite workflow in detail

Use this to train one classifier per marker (e.g. one for CD3 positive/negative, another for CD8 positive/negative) and then combine their results into composite classes. Use it for panels with few markers, or for markers that describe cell state, such as Ki67.

6.1 Create a binary classifier

Menu: Extensions → SP Classify → Binary Classifiers...

1. Click Create... and enter a marker name (e.g. CD3). Characters other than letters, digits, ., _ and - are replaced with _. The name cannot start with . or -.
2. Select the marker in the list and click Open. The dialog closes and the sidebar switches to Binary Mode, with the banner Active binary mode: CD3 in blue.

In binary mode: - The Manual Label Mode class buttons are limited to CD3+ and CD3-. - Pool labels from all images is ticked and cannot be changed. - Settings, sampling, review and metrics work as in multi-class mode. - Predictions are saved per marker in <project>/celltune/binary-image-predictions/<marker>/. Entering or leaving binary mode saves the current predictions and loads the other mode's saved ones. Labels on other images that are not CD3+ or CD3- are skipped.

Repeat train → review → retrain until the Training Metrics F1 stops improving and the predictions look correct on images you did not label. Then click Exit Binary Mode to return to multi-class mode.

Repeat for each marker you want in the composite.

6.2 Composite classification

Menu: Extensions → SP Classify → Composite Classification...

• Markers: one checkbox per trained binary classifier, all ticked by default. Only markers whose classifier has a saved XGBoost model are listed. All / None tick or untick every marker.
• Images: one checkbox per project image, all ticked by default. Buttons: All / None / Current only.
• Prepend current primary classification (colour follows primary): see below.
• Apply: runs the classifiers.

What Apply does: - The open image updates in the viewer immediately. Other selected images are opened, classified and saved. Progress is shown in the dialog's log. - Each cell gets a class made of the marker names in text sort order, each followed by + or -, e.g. CD3+:CD45+:CD8-. In text sort order CD45 comes before CD8. - A marker is + when its binary classifier gives a positive probability of 0.5 or higher, and - otherwise. - The positive probability is the mean of the XGBoost and LightGBM probabilities (XGBoost alone if the marker has no LightGBM model). Each marker uses the batch-correction gains, when correction is on. - If an image lacks a measurement a marker was trained on, that marker is left out of the class names for that image, and the log line starts Skipping '<marker>' not applied.

Prepend current primary classification (off by default): - Unticked: the class name has the marker results only (CD3+:CD8-). QuPath assigns the colour. - Ticked: each cell's existing class is added to the front (Tumour:CD3+:CD8-), and the cell keeps the colour of its existing class. Cells with no class get the marker-only name.

Tick Prepend current primary classification after running a multi-class classifier, so each cell keeps its cell type and gains the marker results.

---