Feature Request: add min_mask_ratio to DeepFeatureExtractor (through SemanticSegmentor.filter_coordinates())

[GeorgeBatch]

• TIA Toolbox version: 1.6.0
• Python version: 3.11
• Operating System: Linux

Description

Existing functionality: When using SlidingWindowPatchExtractor, we can specify min_mask_ratio to filter out patches with little tissue.

Proposed feature: When computing features with DeepFeatureExtractor, we rely on the SemanticSegmentor.filter_coordinates() method, which, as far as I understand based on the definition of sel_func(), needs just one pixel from the patch to be within the mask for the patch to be selected.

tiatoolbox/tiatoolbox/models/engine/semantic_segmentor.py

Lines 634 to 708 in 0ddd2b1

	@staticmethod
	def filter_coordinates(
	mask_reader: VirtualWSIReader,
	bounds: np.ndarray,
	resolution: Resolution | None = None,
	units: Units | None = None,
	) -> np.ndarray:
	"""Indicates which coordinate is valid basing on the mask.
	To use your own approaches, either subclass to overwrite or
	directly assign your own function to this name. In either cases,
	the function must obey the API defined here.
	Args:
	mask_reader (:class:`.VirtualReader`):
	A virtual pyramidal reader of the mask related to the
	WSI from which we want to extract the patches.
	bounds (ndarray and np.int32):
	Coordinates to be checked via the `func`. They must be
	in the same resolution as requested `resolution` and
	`units`. The shape of `coordinates` is (N, K) where N is
	the number of coordinate sets and K is either 2 for
	centroids or 4 for bounding boxes. When using the
	default `func=None`, K should be 4, as we expect the
	`coordinates` to be bounding boxes in `[start_x,
	start_y, end_x, end_y]` format.
	resolution (Resolution):
	Resolution of the requested patch.
	units (Units):
	Units of the requested patch.
	Returns:
	:class:`numpy.ndarray`:
	List of flags to indicate which coordinate is valid.
	Examples:
	>>> # API of function expected to overwrite `filter_coordinates`
	>>> def func(reader, bounds, resolution, units):
	... # as example, only select first bound
	... return np.array([1, 0])
	>>> coords = [[0, 0, 256, 256], [128, 128, 384, 384]]
	>>> segmentor = SemanticSegmentor(model='unet')
	>>> segmentor.filter_coordinates = func
	"""
	if not isinstance(mask_reader, VirtualWSIReader):
	msg = "`mask_reader` should be VirtualWSIReader."
	raise TypeError(msg)
	if not isinstance(bounds, np.ndarray) or not np.issubdtype(
	bounds.dtype,
	np.integer,
	):
	msg = "`coordinates` should be ndarray of integer type."
	raise ValueError(msg)
	mask_real_shape = mask_reader.img.shape[:2]
	mask_resolution_shape = mask_reader.slide_dimensions(
	resolution=resolution,
	units=units,
	)[::-1]
	mask_real_shape = np.array(mask_real_shape)
	mask_resolution_shape = np.array(mask_resolution_shape)
	scale_factor = mask_real_shape / mask_resolution_shape
	scale_factor = scale_factor[0] # what if ratio x != y
	def sel_func(coord: np.ndarray) -> bool:
	"""Accept coord as long as its box contains part of mask."""
	coord_in_real_mask = np.ceil(scale_factor * coord).astype(np.int32)
	start_x, start_y, end_x, end_y = coord_in_real_mask
	roi = mask_reader.img[start_y:end_y, start_x:end_x]
	return np.sum(roi > 0) > 0
	flags = [sel_func(bound) for bound in bounds]
	return np.array(flags)

Benefit: Adding min_mask_ratio can significantly speed up the feature computation by filtering out patches that mostly contain background pixels in advance.