Reference: every setting in the sidebar

Control	Default	What it does
Rounds	500	Maximum boosting rounds (50–1000). With Early stopping ticked, training stops sooner when the model stops improving, so a high value costs little. With it unticked, every round is used; lower the value.
Max depth	6	Maximum tree depth (2–15). Higher values model more complex marker combinations but overfit more easily.
CPU threads	0 (all)	Number of CPU threads training uses; 0 = all cores. Lower it to keep QuPath responsive during training. Remembered between sessions. See §5.3.
Images at once	1	Number of images classified at the same time by Apply to which images... after training (1–8). Each image is loaded whole, so higher values use more memory. Does not affect training.
Model 1	XGBoost	First model type.
Model 2	LightGBM	Second model type. Use a different type from Model 1, so the two models disagree on uncertain cells. XGBoost + XGBoost and LightGBM + LightGBM are refused; Random Forest + Random Forest is allowed.
Pool labels from all images	✅	Train on labelled cells from every project image. Always on in binary mode. Labels whose class is not in the project's class list are skipped; the training log gives the count.
Enable data balancing	✅	Resample the training set to balance the classes. Untick to hide Strategy.
Strategy	SMOTE + Tomek	Resampling method — see the §5.4 table.
Auto-tune hyperparameters	❌	Searches for better model settings. Cost = 2 × Trials × CV folds model fits (200 at the defaults). This can take several hours on a panel with many features. With Train/val metrics ticked, it tunes on the training 80% only. Its scoring macro F1 averages only the classes present in each fold.
Trials	20	Settings combinations tried per model (5–100). Shown only when Auto-tune is ticked.
CV folds	5	Cross-validation folds used to score each combination (2–10). Shown only when Auto-tune is ticked.
Early stopping	✅	Stops adding rounds when the class-weighted log-loss on held-out training cells has not improved for 20 rounds. Below 2,000 labelled cells the held-out cells are 5 cross-validation folds of the training 80%; above, a 10% inner split. XGBoost and LightGBM only; needs at least 20 labelled cells. The 20% validation set is never used.
Train/val metrics	✅	Produces the Training Metrics report. Adds about a third to training time. Untick to train faster without the report. With Auto-tune ticked, it also restricts tuning to the training 80%.
Show top 10 feature importance after training	✅	Opens the SHAP feature-importance plot after training.
Auto-prune features (drop near-constant & redundant)	✅	Before training, removes features that are almost constant or highly correlated with another feature of the same marker. The 5 highest-variance features of each marker are always kept. Runs only when more than 20 features are selected. Image measurements are not changed. See §4.1.
Restrict to features shared with imported data	❌	Train only on features that also exist in the imported ground-truth columns (names matched ignoring case).
Sample current image only	❌	Sample and review cells from the open image only.
Filter by annotation keywords	(blank)	Comma-separated keywords. Only cells inside annotations whose names contain a keyword (any case) are sampled for review.
Apply to which images...	(all)	Choose the project images the trained classifier is applied to. The button label shows how many are selected.
Manual Label Mode	—	Opens the floating labelling toolbar.
Train	—	Starts training. Needs at least 10 labelled cells: on the open image, in imported rows, and (with Pool labels from all images ticked) on other images.
Agreement Confusion Matrix	(disabled)	Shows how often the two models agree, per class. Available after training.
Training Metrics	(disabled)	Per-class precision, recall and F1 on a 20% held-out split. Available after training with Train/val metrics ticked and at least 20 labelled cells.
Feature Importance...	(disabled)	SHAP top features per class. Available after training.
Enter Review Mode	(disabled)	Samples cells where the two models disagree, for you to review. Available once predictions exist.

14.1 Reference: preferences

Under Edit → Preferences → SP Classify.

Preference	Default	What it does
Enable SP Classify extension	✅	Turns the extension off without uninstalling it.
XGBoost histogram bins	0	Leave at 0. See below.
Use batch-corrected values	❌	The same setting as Use batch-corrected values in clustering + ML (streamed, no columns) in the Batch Normalisation dialog. See §19.4.

XGBoost histogram bins. Leave this at 0 (= 256 bins, the most accurate setting). Lower values make XGBoost training faster (128 about 2×, 64 about 2.5×) but change some predictions. To try a lower value, train once at 0 and once at 128, export the cell table each time, and compare the Training Metrics and the class columns. The training log records the value used as XGB max_bin.

---