Classifier

Take an H&E section with tumor nests, stroma and patches of necrosis. You can see the difference at a glance, but no single threshold on one stain separates them. The Classifier learns the difference from examples you draw, then labels the whole image.

1. Classes. In the Classifier main window, add a class, name it and assign a color. Annotation settings (Opacity, Thickness, Close shape, Fill Shape, Fill Alpha) control how your drawings look.
2. Annotations. Click Start drawing, draw over areas that belong to the current class, and click Apply. Start erasing removes mistakes, and double-clicking an annotation in the list centers the image on it for checking.
3. Features. The engine treats the image as blocks of pixels (Object diameter, default 9 × 9) and computes the selected features for each block.
4. Training. A supervised learning step finds how the annotated blocks of each class differ, and produces a trained model that can classify new data.
5. Output. Every block is assigned to a class. Each pixel is overlaid with its class color, and the engine also generates a binary mask for each class.

The Classifier sits in the layer's Pre-Processing section (Add, Operations (Advanced), Classifier), so its output feeds the rest of the layer. It classifies regions, not cells. To give each cell a class, measure the class map under each nucleus with Standard MeasurementsLoading... (global) and its Most Frequent Intensity, or use Assign classes to objectsLoading.... To classify cells by their own appearance, the separate AI Objects Classification engine applies an object model trained in the StrataQuest AI Center.

What the engine works on. Each block of pixels becomes a list of numbers, its features, and classification happens on those numbers, not on the picture. Intensity, Histogram and Percentile features describe the gray levels in the block. HuMoments are, in the textbook form Hu described in 1962, combinations of image moments that stay the same when a shape is moved, scaled or rotated. Haralick features describe texture from the gray-level co-occurrence matrix: how often pairs of gray levels occur side by side.

Filtered images and sizes add context. A block's own pixels say little about the structure it belongs to. That is why features can also be computed on processed versions of the image (Morpho, Average, Gauss, Median), which tie in to morphological operationsLoading... and smoothing filters such as the Gauss filterLoading.... Selected Sizes add proximity areas around each observation, which the reference describes as contextual information. In image processing generally, a stack of filtered images at several scales is a standard way to give a pixel classifier a view of its surroundings.

More features are not always better. In machine learning generally, adding features to a small training set can lower accuracy on new data, because the model fits quirks of the examples. Use Reduced Feature Selection identifies the relevant features during training and computes only those when classifying, which also saves time. The Basic profile is a minimum selection optimized for speed, Advanced is balanced, and All selects every feature.

Annotations determine the outcome. A supervised model can only learn the variation it is shown, and the reference requires a list of classes that covers the variability in your data. In practice that means annotating each class across several regions and slides, including its awkward cases. The reference describes no accuracy score or cross-validation, so checking the map on areas you didn't annotate is your job.

A workable start is Random forest with the Advanced profile, moving to Basic if training is slow. The classifier's own map is not restricted to the ROI. To restrict results to a region of interestLoading..., measure each class in its own layer with Total AreaLoading..., whose detection drops anything outside the ROI.

Source: StrataQuest 8.1 Engines Reference, ch. 7 (Classifier, pp. 366–374), ch. 8 (pp. 374–378), ch. 3 (pp. 8–11), ch. 4 (p. 12); User Manual (pp. 58, 154).

Here is how tumor and stroma could be quantified separately in an IHC or H&E section.

1. In a layer's Pre-Processing, add the Classifier with classes tumor, stroma and necrosis, and annotate each in several regions.
2. Choose Random forest and the Advanced profile, then train.
3. Check the class map against areas you didn't annotate, add annotations where it errs, and train again.
4. For cell counts per class, use the Classifier's "Add class quantification layers" quick action. It adds nuclei detection if none is set, an Assign objects to classesLoading... engine on the class map, a histogramLoading... with a gateLoading... for each class, and a layer per class.
5. For the area of each class, "Add class area quantification layers" imports each class mask into its own layer and measures it with Total Area and Standard Measurements.

The two routes answer different questions. The first measures how many cells, and of which kind, sit in tumor versus stroma. The second measures how much of the section each compartment occupies, which is the denominator for densities.