nucleus_detector¶

tiatoolbox . models . engine . nucleus_detector

Nucleus Detection Engine for Digital Pathology (WSIs and patches).

This module implements the NucleusDetector class which extends SemanticSegmentor to perform instance-level nucleus detection on histology images. It supports patch-mode and whole slide image (WSI) workflows using TIAToolbox or custom PyTorch models, and provides utilities for parallel post-processing (centroid extraction, thresholding), merging detections across patches, and exporting results in multiple formats (in-memory dict, Zarr, AnnotationStore).

Classes¶

NucleusDetectorRunParams: TypedDict specifying runtime configuration keys for detection.
NucleusDetector: Core engine for nucleus detection on image patches or WSIs.

Examples:¶

>>> from tiatoolbox.models.engine.nucleus_detector import NucleusDetector
>>> detector = NucleusDetector(model="mapde-conic")
>>> # WSI workflow: save to AnnotationStore (.db)
>>> out = detector.run(
...     images=[pathlib.Path("example_wsi.tiff")],
...     patch_mode=False,
...     device="cuda",
...     save_dir=pathlib.Path("output_directory/"),
...     overwrite=True,
...     output_type="annotationstore",
...     class_dict={0: "nucleus"},
...     auto_get_mask=True,
...     memory_threshold=80,
... )
>>> # Patch workflow: return in-memory detections
>>> patches = [np.ndarray, np.ndarray]  # NHWC
>>> out = detector.run(patches, patch_mode=True, output_type="dict")

Notes:¶

Outputs can be returned as Python dictionaries, saved as Zarr groups, or converted to AnnotationStore (.db).
Post-processing uses tile rechunking and halo padding to facilitate centroid extraction near chunk boundaries.

Functions

`save_detection_arrays_to_qupath_json`	Write nucleus detection arrays to QuPath JSON.
`save_detection_arrays_to_store`	Write nucleus detection arrays to an SQLite-backed AnnotationStore.

Classes

`NucleusDetector`	Nucleus detection engine for digital histology images.
`NucleusDetectorRunParams`	Runtime parameters for configuring the NucleusDetector.run() method.