|
|
""" |
|
|
Copyright 2021, Dana-Farber Cancer Institute and Weill Cornell Medicine |
|
|
License: GNU GPL 2.0 |
|
|
""" |
|
|
|
|
|
import numpy as np |
|
|
|
|
|
|
|
|
def extract_tiles(arr, tile_size, stride=None): |
|
|
""" |
|
|
Extract tiles from an array. Allows user to specify stride between tiles. |
|
|
Based on original implementation in ``sklearn.feature_extraction.image_ref._extract_patches`` |
|
|
Incomplete tiles on the edge are dropped (TO DO: fix this (zero padding?)) |
|
|
|
|
|
Args: |
|
|
arr (np.ndarray): input array. Must be 3 dimensional (H, W, n_channels) |
|
|
tile_size (int): Dimension of extracted tiles. Each tile will be shape (tile_size, tile_size, n_channels) |
|
|
stride (int, optional): Stride length between tiles. |
|
|
If ``None``, uses ``stride = tile_size`` for non-overlapping tiles. Defaults to None |
|
|
|
|
|
Returns: |
|
|
np.ndarray: Array of extracted tiles of shape `(n_tiles, tile_size, tile_size, n_channels)` |
|
|
""" |
|
|
assert arr.ndim == 3, f"Number of input dimensions {arr.ndim} must be 3" |
|
|
if stride is None: |
|
|
stride = tile_size |
|
|
i, j, n_channels = arr.shape |
|
|
if (i - tile_size) % stride != 0 or (j - tile_size) % stride != 0: |
|
|
raise NotImplementedError( |
|
|
f"Array of shape {arr.shape} is not perfectly tiled by tiles of size {tile_size} and stride {stride}." |
|
|
) |
|
|
patch_strides = arr.strides |
|
|
patch_shape = (tile_size, tile_size, n_channels) |
|
|
extraction_step = (stride, stride, 1) |
|
|
slices = tuple(slice(None, None, st) for st in extraction_step) |
|
|
indexing_strides = arr[slices].strides |
|
|
patch_indices_shape = ( |
|
|
(np.array(arr.shape) - np.array(patch_shape)) // np.array(extraction_step) |
|
|
) + 1 |
|
|
shape = tuple(list(patch_indices_shape) + list(patch_shape)) |
|
|
strides = tuple(list(indexing_strides) + list(patch_strides)) |
|
|
tiles = np.lib.stride_tricks.as_strided(arr, shape=shape, strides=strides) |
|
|
|
|
|
tiles = tiles.squeeze(axis=2) |
|
|
tiles = tiles.reshape(-1, *tiles.shape[2:]) |
|
|
return tiles |
|
|
|
|
|
|
|
|
def extract_tiles_with_mask(arr, mask, tile_size, stride=None, threshold=0.5): |
|
|
""" |
|
|
Generate tiles and only keep tiles that overlap with the |
|
|
masks above some threshold. |
|
|
|
|
|
Args: |
|
|
arr (np.ndarray): input array. Must be 3 dimensional (H, W, n_channels) |
|
|
mask (np.ndarray): array of masks. Must be 3 dimensional (H, W, n_masks). |
|
|
tile_size (int): Dimension of extracted tiles. Each tile will be shape (tile_size, tile_size, n_channels) |
|
|
stride (int, optional): Stride length between tiles. |
|
|
If ``None``, uses ``stride = tile_size`` for non-overlapping tiles. Defaults to None |
|
|
threshold (float): for each tile, all values in the corresponding mask region will be averaged. |
|
|
`threshold` is the cutoff value, above which the tile will be kept and below which the tile |
|
|
will be discarded. Defaults to 0.5 |
|
|
|
|
|
Returns: |
|
|
np.ndarray: Array of extracted tiles of shape `(n_tiles, tile_size, tile_size, n_channels)` |
|
|
""" |
|
|
|
|
|
assert isinstance( |
|
|
mask, np.ndarray |
|
|
), f"Input mask type {type(mask)} must be a numpy array" |
|
|
assert isinstance( |
|
|
arr, np.ndarray |
|
|
), f"input array type {type(arr)} must be a numpy array" |
|
|
assert arr.ndim == 3, f"array of shape {arr.shape} must be 3 dimensional" |
|
|
assert mask.ndim == 3, f"mask of shape {mask.shape} must be 3 dimensional" |
|
|
assert ( |
|
|
arr.shape[0:2] == mask.shape[0:2] |
|
|
), f"Dims of input image_ref {arr.shape} and mask {mask.shape} must match" |
|
|
|
|
|
arr_tiles = extract_tiles(arr, tile_size=tile_size, stride=stride) |
|
|
mask_tiles = extract_tiles(mask, tile_size=tile_size, stride=stride) |
|
|
|
|
|
tile_mask_means = mask_tiles.mean(axis=tuple(range(1, mask_tiles.ndim))) |
|
|
|
|
|
out = arr_tiles[tile_mask_means >= threshold, ...] |
|
|
return out |
|
|
|
|
|
|
|
|
|
|
|
|
