def get_bounding_box_around_mask(self, color):
# if so, then determine how much to resize things by for the bounding box
slide = ops.open_slide(self.svsPath)
dz = deepzoom.DeepZoomGenerator(slide,
tile_size=self.patch_size,
overlap=0,
limit_bounds=True)
max_mag = int(slide.properties['openslide.objective-power'])
resize_factor = float(self.mag / max_mag)
# determine the original bounding box from the annotation
regions = self.pxml.get_all_regions(color)
all_x = np.array([])
all_y = np.array([])
for idx, each_region in enumerate(regions):
for each_vertex in each_region.find_all('vertex'):
all_x = np.append(all_x, int(float(each_vertex.get('x'))))
all_y = np.append(all_y, int(float(each_vertex.get('y'))))
all_x = all_x * resize_factor
all_y = all_y * resize_factor
# apply resize factor
min_x = int(np.min(all_x))
max_x = int(np.max(all_x))
min_y = int(np.min(all_y))
max_y = int(np.max(all_y))
return (min_x, min_y, max_x, max_y)
def extract_roi(self):
# get useful information
start_col, start_row, end_col, end_row = self.find_tile_in_dz()
slide = ops.open_slide(self.svsPath)
dz = deepzoom.DeepZoomGenerator(slide,
tile_size=self.patch_size,
overlap=0,
limit_bounds=True)
level = self.dict_mag_level[self.mag]
print("level: ", level)
#initialize empty array the size of the ROI we need
w = (end_col - start_col) * self.patch_size
h = (end_row - start_row) * self.patch_size
io = np.zeros((h, w, 3), dtype='uint8')
# not sure if it should by y,x or x,y but either all of the rows and columns needs to be flipped
for i in range(start_col, end_col):
for j in range(start_row, end_row):
patch = dz.get_tile(level, (i, j))
x_start = (i - start_col) * self.patch_size
x_end = x_start + self.patch_size
y_start = (j - start_row) * self.patch_size
y_end = y_start + self.patch_size
io[y_start:y_end, x_start:x_end, :] = patch
return io
def get_slide_tile_information(self):
slide = ops.open_slide(self.svsPath)
dz = deepzoom.DeepZoomGenerator(slide,
tile_size=self.patch_size,
overlap=0,
limit_bounds=True)
level = self.dict_mag_level[mag]
w, h = dz.level_dimensions[level]
n, m = dz.level_tiles[level]
return (n, m, w, h)
def svs2dask_array(svs_file,
tile_size=1000,
overlap=0,
remove_last=True,
allow_unknown_chunksizes=False):
"""Convert SVS, TIF or TIFF to dask array.
Parameters
----------
svs_file:str
Image file.
tile_size:int
Size of chunk to be read in.
overlap:int
Do not modify, overlap between neighboring tiles.
remove_last:bool
Remove last tile because it has a custom size.
allow_unknown_chunksizes: bool
Allow different chunk sizes, more flexible, but slowdown.
Returns
-------
dask.array
Dask Array.
>>> arr=svs2dask_array(svs_file, tile_size=1000, overlap=0, remove_last=True, allow_unknown_chunksizes=False)
>>> arr2=arr.compute()
>>> arr3=to_pil(cv2.resize(arr2, dsize=(1440,700), interpolation=cv2.INTER_CUBIC))
>>> arr3.save(test_image_name)"""
img = openslide.open_slide(svs_file)
gen = deepzoom.DeepZoomGenerator(img,
tile_size=tile_size,
overlap=overlap,
limit_bounds=True)
max_level = len(gen.level_dimensions) - 1
n_tiles_x, n_tiles_y = gen.level_tiles[max_level]
get_tile = lambda i, j: np.array(gen.get_tile(max_level,
(i, j))).transpose((1, 0, 2))
sample_tile = get_tile(0, 0)
sample_tile_shape = sample_tile.shape
dask_get_tile = dask.delayed(get_tile, pure=True)
arr = da.concatenate([
da.concatenate([
da.from_delayed(dask_get_tile(i, j), sample_tile_shape, np.uint8)
for j in range(n_tiles_y - (0 if not remove_last else 1))
],
allow_unknown_chunksizes=allow_unknown_chunksizes,
axis=1)
for i in range(n_tiles_x - (0 if not remove_last else 1))
],
allow_unknown_chunksizes=allow_unknown_chunksizes
) #.transpose([1,0,2])
return arr
def buildDictionary(self):
slide = ops.open_slide(self.svsPath)
dz = deepzoom.DeepZoomGenerator(slide,
tile_size=self.patch_size,
overlap=0,
limit_bounds=True)
levels = dz.level_count
max_mag = int(slide.properties['openslide.objective-power'])
counter = 1
dict_level_mag_correspondence = {}
for i in reversed(range(0, levels)):
dict_level_mag_correspondence[max_mag / counter] = i
counter = counter * 2
return dict_level_mag_correspondence
def svs2dask_array(svs_file, tile_size=1000, overlap=0, remove_last=True, allow_unknown_chunksizes=False, transpose=False): """Convert SVS, TIF or TIFF to dask array. Parameters ---------- svs_file : str Image file. tile_size : int Size of chunk to be read in. overlap : int Do not modify, overlap between neighboring tiles. remove_last : bool Remove last tile because it has a custom size. allow_unknown_chunksizes : bool Allow different chunk sizes, more flexible, but slowdown. Returns ------- arr : dask.array.Array A Dask Array representing the contents of the image file. >>> arr = svs2dask_array(svs_file, tile_size=1000, overlap=0, remove_last=True, allow_unknown_chunksizes=False) >>> arr2 = arr.compute() >>> arr3 = to_pil(cv2.resize(arr2, dsize=(1440, 700), interpolation=cv2.INTER_CUBIC)) >>> arr3.save(test_image_name) """ # https://github.com/jlevy44/PathFlowAI/blob/master/pathflowai/utils.py img = openslide.open_slide(svs_file) if type(img) is openslide.OpenSlide: gen = deepzoom.DeepZoomGenerator( img, tile_size=tile_size, overlap=overlap, limit_bounds=True) max_level = len(gen.level_dimensions) - 1 n_tiles_x, n_tiles_y = gen.level_tiles[max_level] @dask.delayed(pure=True) def get_tile(level, column, row): tile = gen.get_tile(level, (column, row)) return np.array(tile).transpose((1, 0, 2)) sample_tile_shape = get_tile(max_level, 0, 0).shape.compute() rows = range(n_tiles_y - (0 if not remove_last else 1)) cols = range(n_tiles_x - (0 if not remove_last else 1)) arr = da.concatenate([da.concatenate([da.from_delayed(get_tile(max_level, col, row), sample_tile_shape, np.uint8) for row in rows], allow_unknown_chunksizes=allow_unknown_chunksizes, axis=1) for col in cols], allow_unknown_chunksizes=allow_unknown_chunksizes) if transpose: arr=arr.transpose([1, 0, 2]) return arr else: # img is instance of openslide.ImageSlide return dask_image.imread.imread(svs_file)
def get_preview(slidename, classifier, level, size, slide_folder, neg=0):
result = []
slidename = os.path.basename(slidename)
slidepath = os.path.join(slide_folder, slidename)
slide = OpenSlide(slidepath)
slide_dz = deepzoom.DeepZoomGenerator(
slide, tile_size=(size - 2), overlap=1)
tiles = slide_dz.level_tiles[level]
preview = numpy.zeros(tiles)
for x in classifier:
im_x = int(x[1])
im_y = int(x[2])
if x[4] == neg and x[5] == neg:
preview[im_x][im_y] = 1
else:
cluster = x[4]
preview[im_x][im_y] = cluster + 2
result.extend((slidename, preview))
return result
def get_probability_images(slidename, classifier, slide_folder, features=30):
result = []
slidepath = os.path.join(slide_folder, slidename)
slide = OpenSlide(slidepath)
slide_dz = deepzoom.DeepZoomGenerator(slide,
tile_size=(224 - 2),
overlap=1)
tiles = slide_dz.level_tiles[16]
preview = numpy.zeros((tiles[0], tiles[1], features + 1))
for x in classifier:
im_x = int(x[1])
im_y = int(x[2])
if x[3] == 0:
preview[im_x][im_y][0] = 1
else:
for f in range(features):
preview[im_x][im_y][f + 1] = x[4 + f]
image_name = os.path.join(outpath,
'{}#{}features.npy'.format(slidename, features))
numpy.save(image_name, preview)
return image_name
def get(self, path):
with self._lock:
if path in self._cache:
# Move to end of LRU
slide = self._cache.pop(path)
self._cache[path] = slide
return slide
osr = openslide.OpenSlide(path)
slide = deepzoom.DeepZoomGenerator(osr, **self.dz_opts)
try:
mpp_x = osr.properties[openslide.PROPERTY_NAME_MPP_X]
mpp_y = osr.properties[openslide.PROPERTY_NAME_MPP_Y]
slide.mpp = (float(mpp_x) + float(mpp_y)) / 2
except (KeyError, ValueError):
slide.mpp = 0
with self._lock:
if path not in self._cache:
if len(self._cache) == self.cache_size:
self._cache.popitem(last=False)
self._cache[path] = slide
return slide
def extract_patches(img_id, msk_id, level=1, sz=128):
"""
This function splits each slide into patches and save them onto your
computer. Each patch is saved based on its class (stroma, benign, pattern 3,
4, and 5). Patches from the same slide is under the same slide id folder. A
csv file containing each patch's filename and class is generated. A csv file
containing slides IDs of extracted patches is also generated.
"""
patches = [] # Store patch filename and class
# Here 128 doesn't work for some reasons, so 126 is chosen. No overlapping
# between patches. Lower this number can generating patches with overlapping
# areas.
tile_size = 126
overlap = int(
(sz - tile_size) / 2) # Size of overlapping areas between patches
# Output csv path for patch filename and class
csv_path = "./prostate-cancer-grade-assessment/patches_level1_" + str(
sz) + ".csv"
for i in trange(len(img_id)):
# Read slides
im = openslide.OpenSlide(images_path + img_id[i])
dpz = deepzoom.DeepZoomGenerator(im,
tile_size=tile_size,
overlap=overlap,
limit_bounds=False)
width = dpz.level_tiles[dpz.level_count - 3][0]
height = dpz.level_tiles[dpz.level_count - 3][1]
offset = int(np.ceil(overlap / tile_size))
# Store slide tiles into a dictionary
tiles1 = {}
count = 0
for j in range(offset,
dpz.level_tiles[dpz.level_count - 3][1] - 1 - offset):
for k in range(
offset,
dpz.level_tiles[dpz.level_count - 3][0] - 1 - offset):
tiles1[count] = np.asarray(
dpz.get_tile(dpz.level_count - 3, (k, j)))
count += 1
im.close()
# Read masks
im = openslide.OpenSlide(masks_path + msk_id[i])
dpz = deepzoom.DeepZoomGenerator(im,
tile_size=tile_size,
overlap=overlap,
limit_bounds=False)
width = dpz.level_tiles[dpz.level_count - 3][0]
height = dpz.level_tiles[dpz.level_count - 3][1]
offset = int(np.ceil(overlap / tile_size))
# Store mask tiles into a dictionary
tiles2 = {}
count = 0
for j in range(offset,
dpz.level_tiles[dpz.level_count - 3][1] - 1 - offset):
for k in range(
offset,
dpz.level_tiles[dpz.level_count - 3][0] - 1 - offset):
tiles2[count] = np.asarray(
dpz.get_tile(dpz.level_count - 3, (k, j)))
count += 1
im.close()
if len(tiles1) > 0:
# Rank slide tiles by counting color pixels on mask tiles
tiles2 = {
k: v
for k, v in sorted(tiles2.items(),
key=lambda item: item[1].sum(),
reverse=True)
}
iteration = 0
# Score mask tiles
for tile_idx in tiles2.keys():
score, relative_score = grade(tiles2[tile_idx])
# Find class
# Patches contain more than 1 color are not used
tile_score = -1
if np.sum(score) == 1:
if np.sort(relative_score)[:3].sum() == 0:
if score[np.argmax(relative_score) + 1] > 0.4:
tile_score = np.argmax(relative_score) + 2
elif score[0] > 0.7:
tile_score = 1
if tile_score >= 1:
if if_clean(tiles1[tile_idx]):
# Save patches
im = Image.fromarray(tiles1[tile_idx])
temp1 = "./prostate-cancer-grade-assessment/patches_level1_" + \
str(sz) + "/"
temp2 = img_id[i].split(".")[0] + "/"
temp3 = str(tile_score) + "/"
if not os.path.isdir(temp1):
os.mkdir(temp1)
if not os.path.isdir(temp1 + temp2):
os.mkdir(temp1 + temp2)
if not os.path.isdir(temp1 + temp2 + temp3):
os.mkdir(temp1 + temp2 + temp3)
# Filename example: /#slideID/#classNumber/#patchFilename.tiff
im.save(temp1 + temp2 + temp3 + str(iteration) +
img_id[i])
# Append patch filenamd and class to csv
patches.append(
[str(iteration) + img_id[i],
str(tile_score)])
iteration += 1
# Save csv file containing patch filename and class
np.savetxt(csv_path, patches, delimiter=",", fmt="%s")
# Save csv file containing slide IDs of extracted patches
folders = os.listdir(temp1)
csv_content = []
for folder in folders:
if os.path.isdir(folder):
csv_content.append(folder)
np.savetxt(patches_level1_128_slideFolders_path,
csv_content,
delimiter=",",
fmt="%s")
def __create_tiles(self, mask, bg_color):
"""Create tiles given a PySlide and a mask.
Arguments:
mask: PIL Image containing the mask for the slide.
bg_color: Numpy array indicating the color used for the background in the mask.
"""
ts = time.time()
# Create folder for the patches
if self.input_slide.save_patches:
self.input_slide._create_tile_folder()
# Initialize deep zoom generator for the slide
image_dims = self.input_slide.slide.dimensions
dzg = deepzoom.DeepZoomGenerator(self.input_slide.slide, tile_size=self.input_slide.patch_size, overlap=0)
# Find the deep zoom level corresponding to the
# requested downsampling factor
dzg_levels = [2**i for i in range(0, dzg.level_count)][::-1]
dzg_selectedlevel_idx = dzg_levels.index(self.input_slide.output_downsample)
dzg_selectedlevel_dims = dzg.level_dimensions[dzg_selectedlevel_idx]
dzg_selectedlevel_maxtilecoords = dzg.level_tiles[dzg_selectedlevel_idx]
dzg_real_downscaling = np.divide(image_dims, dzg.level_dimensions)[:, 0][dzg_selectedlevel_idx]
n_tiles = np.prod(dzg_selectedlevel_maxtilecoords)
digits_padding = len(str(n_tiles))
# Calculate patch size in the mask
mask_patch_size = int(np.ceil(self.input_slide.patch_size * (self.input_slide.output_downsample/self.input_slide.mask_downsample)))
# Deep zoom generator for the mask
dzgmask = deepzoom.DeepZoomGenerator(openslide.ImageSlide(mask), tile_size=mask_patch_size, overlap=0)
dzgmask_dims = dzgmask.level_dimensions[dzgmask.level_count - 1]
dzgmask_maxtilecoords = dzgmask.level_tiles[dzgmask.level_count - 1]
dzgmask_ntiles = np.prod(dzgmask_maxtilecoords)
# If needed, generate an image to store tile-crossed output, at the requested tilecross downsample level
if self.input_slide.save_tilecrossed_image:
# Get a downsampled numpy array for the image
tilecrossed_img = utility_functions.downsample_image(self.input_slide.slide,
self.input_slide.tilecross_downsample, mode="numpy")[0]
# Calculate patch size in the mask
tilecross_patchsize = int(np.ceil(self.input_slide.patch_size * (self.input_slide.output_downsample/self.input_slide.tilecross_downsample)))
# Draw the grid at the scaled patchsize
x_shift, y_shift = tilecross_patchsize, tilecross_patchsize
gcol = [255, 0, 0]
tilecrossed_img[:, ::y_shift, :] = gcol
tilecrossed_img[::x_shift, :, :] = gcol
# Convert numpy array to PIL image
tilecrossed_img = Image.fromarray(tilecrossed_img, mode="RGB")
# Create object to draw the crosses for each tile
draw = ImageDraw.Draw(tilecrossed_img)
# Counters for iterating through the tile-crossed image tiles
tc_w = 0
tc_h = 0
# Debug information
logging.debug("** Original image information **")
logging.debug("-Dimensions: " + str(image_dims))
logging.debug("** Mask information **")
logging.debug("-Mask downscaling factor: " + str(self.input_slide.mask_downsample))
logging.debug("-Pixel dimensions: " + str(dzgmask_dims))
logging.debug("-Calculated patch size: " + str(mask_patch_size))
logging.debug("-Max tile coordinates: " + str(dzgmask_maxtilecoords))
logging.debug("-Number of tiles: " + str(dzgmask_ntiles))
logging.debug("** Output image information **")
logging.debug("Requested " + str(self.input_slide.output_downsample) + "x downsampling for output.")
logging.debug("** Properties of selected deep zoom level **")
logging.debug("-Real downscaling factor: " + str(dzg_real_downscaling))
logging.debug("-Pixel dimensions: " + str(dzg_selectedlevel_dims))
logging.debug("-Selected patch size: " + str(self.input_slide.patch_size))
logging.debug("-Max tile coordinates: " + str(dzg_selectedlevel_maxtilecoords))
logging.debug("-Number of tiles: " + str(n_tiles))
logging.info("== Selecting tiles ==")
if dzgmask_maxtilecoords != dzg_selectedlevel_maxtilecoords:
logging.info("Rounding error creates extra patches at the side(s) of the image.")
grid_coord = (min(dzgmask_maxtilecoords[0], dzg_selectedlevel_maxtilecoords[0]),
min(dzgmask_maxtilecoords[1], dzg_selectedlevel_maxtilecoords[1]))
logging.info("Ignoring the image border. Maximum tile coordinates: " + str(grid_coord))
n_tiles = grid_coord[0] * grid_coord[1]
else:
grid_coord = dzg_selectedlevel_maxtilecoords
# Counters
preds = [0] * n_tiles
row, col, i = 0, 0, 0
tile_names = []
tile_dims_w = []
tile_dims_h = []
tile_rows = []
tile_cols = []
# Evaluate tiles using the selector function
while row < grid_coord[1]:
# Extract the tile from the mask (the last level is used
# since the mask is already rescaled)
mask_tile = dzgmask.get_tile(dzgmask.level_count - 1, (col, row))
# Tile converted to BGR
mask_tile = np.array(mask_tile)
# Predict if the tile will be kept (1) or not (0)
preds[i] = utility_functions.selector(mask_tile, self.input_slide.thres, bg_color, self.input_slide.method)
# Save patches if requested
if self.input_slide.save_patches:
tile = dzg.get_tile(dzg_selectedlevel_idx, (col, row))
# If we need square patches only, we set the prediction to zero if the tile is not square
if not self.input_slide.save_nonsquare:
if tile.size[0] != tile.size[1]:
preds[i] = 0
# Prepare metadata
tile_names.append(self.input_slide.sample_id + "_" + str(i).zfill(digits_padding))
tile_dims_w.append(tile.size[0])
tile_dims_h.append(tile.size[1])
tile_rows.append(row)
tile_cols.append(col)
# Save tile
imgtile_out = self.input_slide.tile_folder + tile_names[i] + "." + self.input_slide.format
if self.input_slide.save_blank:
tile.save(imgtile_out)
else:
if preds[i] == 1:
tile.save(imgtile_out)
# Draw cross over corresponding patch section on tilecrossed image
if self.input_slide.save_tilecrossed_image:
start_w = col * (tilecross_patchsize)
start_h = row * (tilecross_patchsize)
# If we reach the edge of the image, we only can draw until the edge pixel
if (start_w + tilecross_patchsize) >= tilecrossed_img.size[0]:
cl_w = tilecrossed_img.size[0] - start_w
else:
cl_w = tilecross_patchsize
if (start_h + tilecross_patchsize) >= tilecrossed_img.size[1]:
cl_h = tilecrossed_img.size[1] - start_h
else:
cl_h = tilecross_patchsize
# Draw the cross only if the tile has to be kept
if preds[i] == 1:
# From top left to bottom right
draw.line([(start_w, start_h), (start_w + cl_w, start_h + cl_h)], fill=(0, 0, 255), width=3)
# From bottom left to top right
draw.line([(start_w, start_h + cl_h), (start_w + cl_w, start_h)], fill=(0, 0, 255), width=3)
# Jump to the next tilecross tile
tc_w = tc_w + tilecross_patchsize + 1
# Jump to the next column tile
col += 1
# If we reach the right edge of the image, jump to the next row
if col == grid_coord[0]:
col = 0
row += 1
if self.input_slide.save_tilecrossed_image:
tc_w = 0
tc_h = tc_h + tilecross_patchsize + 1
# Increase counter for metadata
i += 1
# Saving tilecrossed image
if self.input_slide.save_tilecrossed_image:
tilecrossed_outpath = self.input_slide.img_outpath + "/tilecrossed_" + self.input_slide.sample_id + "." + self.input_slide.format
tilecrossed_img.save(tilecrossed_outpath)
# Save predictions for each tile
if self.input_slide.save_patches:
patch_results = []
patch_results.extend(list(zip(tile_names, tile_dims_w, tile_dims_h, preds, tile_rows, tile_cols)))
patch_results_df = pd.DataFrame.from_records(patch_results, columns=["Tile", "Width", "Height", "Keep", "Row", "Column"])
patch_results_df.to_csv(self.input_slide.img_outpath + "tile_selection.tsv", index=False, sep="\t")
# Finishing
te = time.time()
logging.debug("Elapsed time: " + str(round(te - ts, ndigits = 3)) + "s")
if self.input_slide.save_blank:
logging.debug("Selected " + str(len(preds)) + " tiles")
else:
logging.debug("Selected " + str(sum(preds)) + " tiles")
im = openslide.OpenSlide(images_path +
train_image_id[new_train_idx])
im2 = im.read_region(location=(0, 0), level=level, \
size=(im.level_dimensions[level][0],\
im.level_dimensions[level][1]))
width = im.level_dimensions[level][0]
height = im.level_dimensions[level][1]
# 0:3 because the 4th element, alpha, is not being used
data = np.array(im2)[:, :, 0:3]
ax[0].imshow(data)
# Split into patches
dpz = deepzoom.DeepZoomGenerator(im,
tile_size=tile_size,
overlap=overlap,
limit_bounds=False)
width = dpz.level_tiles[dpz.level_count - 3][0]
height = dpz.level_tiles[dpz.level_count - 3][1]
offset = int(np.ceil(overlap / tile_size))
tiles1 = []
for j in range(offset,
dpz.level_tiles[dpz.level_count - 3][1] - offset):
for i in range(
offset,
dpz.level_tiles[dpz.level_count - 3][0] - offset):
tiles1.append(
np.asarray(dpz.get_tile(dpz.level_count - 3, (i, j))))
tiles1 = np.stack(tiles1, axis=0)
def predict_wsi(model, global_fixed, slide_path):
# size_g, size_p = (244, 244), (244, 244)
# size_g, size_p = (1008, 1008), (1008, 1008)
# n_class = 2
# sub_batch_size = 1
def predict(image_as_tensor,
size_g=(244, 244),
size_p=(244, 244),
n_class=2):
images_glb = resize(image_as_tensor, size_g)
scores = [
np.zeros((1, n_class, image_as_tensor[i].size[1],
image_as_tensor[i].size[0]))
for i in range(len(image_as_tensor))
]
images_glb = images_transform(images_glb)
patches, coordinates, templates, sizes, ratios = global2patch(
image_as_tensor, size_p)
predicted_ensembles = [
np.zeros((len(coordinates[i]), n_class, size_p[0], size_p[1]))
for i in range(len(image_as_tensor))
]
for i in range(len(image_as_tensor)):
j = 0
while j < len(coordinates[i]):
patches_var = images_transform(patches[i][j:j +
1]) # b, c, h, w
fm_patches, _ = model.module.collect_local_fm(
images_glb[i:i + 1],
patches_var,
ratios[i],
coordinates[i],
[j, j + 1],
len(image_as_tensor),
global_model=global_fixed,
template=templates[i],
n_patch_all=len(coordinates[i]),
)
j += 1
_, fm_global = model.forward(images_glb,
None,
None,
None,
mode=PhaseMode.GlobalFromLocal)
for i in range(len(image_as_tensor)):
j = 0
# while j < n ** 2:
while j < len(coordinates[i]):
fl = fm_patches[i][j:j + 1].cuda()
fg = crop_global(fm_global[i:i + 1], coordinates[i][j:j + 1],
ratios[i])[0]
fg = F.interpolate(fg, size=fl.size()[2:], mode="bilinear")
output_ensembles = model.module.ensemble(
fl, fg) # include cordinates
# output_ensembles = F.interpolate(model.module.ensemble(fl, fg), size_p, **model.module._up_kwargs)
# ensemble predictions
predicted_ensembles[i][j:j + output_ensembles.size()[0]] += (
F.interpolate(
output_ensembles,
size=size_p,
mode="nearest",
).data.cpu().numpy())
j += 1
scores[i] += np.rot90(
np.array(
patch2global(
predicted_ensembles[i:i + 1],
n_class,
sizes[i:i + 1],
coordinates[i:i + 1],
size_p,
)),
k=0,
axes=(3, 2),
)
return [score.argmax(1)[0] for score in scores]
slide = OpenSlide(slide_path)
w, h = slide.level_dimensions[2]
img = slide.read_region((0, 0), 2, (w, h))
slide.close()
img.convert('RGB').save('/tmp/temp.jpg')
slide = ImageSlide('/tmp/temp.jpg')
dz = deepzoom.DeepZoomGenerator(slide, tile_size=1024, overlap=0)
cols, rows = dz.level_tiles[-1]
out = np.zeros((rows * 1024, cols * 1024), dtype=np.uint8)
for row in range(rows):
for col in range(cols):
tile = dz.get_tile(dz.level_count - 1, (col, row)) # col, row
tile_coors = dz.get_tile_coordinates(dz.level_count - 1,
(col, row))
left, top = tile_coors[0]
t_w, t_h = tile_coors[2]
if tile.size != (1024, 1024):
tile = add_extra_pixels(tile, expected_shape=(1024, 1024))
tile = np.array(tile)
processed = apply_filters(tile)
pred = predict([Image.fromarray(processed)])
pil_pred = pred[0].astype(np.uint8)
newmask = remove_mask_overlay_background(processed, pil_pred)
# applied_mask = apply_mask(tile, newmask)
applied_mask = newmask
out[top:top + t_h, left:left + t_w] = applied_mask[:t_h, :t_w]
return out[:h, :w]
def makeSplitData(folderName, fileNameHere, dirName):
print(fileNameHere)
splitNum = 0
curFolder = dirName + folderName + '/'
svsFile = fileNameHere + '.svs'
slide = OpenSlide(curFolder + svsFile)
magVal = slide.properties['aperio.AppMag']
mppVal = slide.properties['aperio.MPP']
zoomed = deepzoom.DeepZoomGenerator(slide,
tile_size=360,
overlap=0,
limit_bounds=False)
levelCount = int(zoomed.level_count)
maxD = slide.level_dimensions[0]
maxY = maxD[1]
maxX = maxD[0]
levelTiles = zoomed.level_tiles
if magVal == '40':
targetLevel = levelCount - 2
totTiles = levelTiles[targetLevel]
maxY = int(maxD[1] / 2)
maxX = int(maxD[0] / 2)
elif magVal == '20' and mppVal[:3] == '0.2':
targetLevel = levelCount - 2
totTiles = levelTiles[targetLevel]
maxY = int(maxD[1] / 2)
maxX = int(maxD[0] / 2)
elif magVal == '20':
targetLevel = levelCount - 1
totTiles = levelTiles[targetLevel]
maxY = int(maxD[1])
maxX = int(maxD[0])
else:
print('mag is not 20 or 40!!!')
return
train_batch = []
train_tile = []
train_ij = []
batch_num = 0
for i in range(totTiles[0] - 1):
for j in range(totTiles[1] - 1):
if i % 10 == 0 and j % 10 == 0:
print('(' + str(i) + ',' + str(j) + ')')
tile = zoomed.get_tile(targetLevel, (i, j))
train_tile.append(tile)
tileStartX = i * 360
tileEndX = (i + 1) * 360
tileStartY = j * 360
tileEndY = (j + 1) * 360
batch_num += 1
train_batch.append(np.array(tile))
train_ij.append((i, j))
if batch_num == 256 or (i == (totTiles[0] - 2)
and j == (totTiles[1] - 2)):
result = sess.run((cp),
feed_dict={
cx: train_batch,
train_mode: False
})
for k in range(len(result)):
if result[k][1] > 0.5:
train_tile[k].save(curFolder + '/TX/' + fileNameHere +
'-' + str(train_ij[k][0]) + '-' +
str(train_ij[k][1]) + '.jpg')
splitNum += 1
else:
train_tile[k].save(curFolder + '/NT/' + fileNameHere +
'-' + str(train_ij[k][0]) + '-' +
str(train_ij[k][1]) + '.jpg')
splitNum += 1
batch_num = 0
train_batch = []
train_ij = []
train_tile = []
return splitNum
def setup():
slide = OpenSlide('HCI/slide-webserver/slideserver/static/testslide.svs')
app.deepzoom = deepzoom.DeepZoomGenerator(slide)
def get_patches(slidepath, outpath, level=10, tissue_ratio=0.25, size=256):
"""
Function that divides a slide into patches with different resolution. The
patches are saved inside a folder with the slide name, and have the format
{slide_name}#{patch_number}-level{}-{x}-{y}.jpg. It also saves a preview
for each slide under the format slidename.png
Arguments:
- slidepath: str, path to the image to patchify
- outpath: str, path to the folder in which a new folder will be
created, where the patches will be saved. This folder has the same
name as the image to patchify
- level: int, level in which image is patchified. The bigger the level,
the higher the number of patches and the resolution of the images.
Default = 16
- tissue_ratio: float, minimum surface of tissue tile to be considered.
Default = 0.25
- size: int, side number of pixels (n pixels size*size). Default = 256
Returns:
- n: int, number of patches
- outpath: str, path to folder where the patches are saved
"""
# Opens the slide with OpenSlide
slide = OpenSlide(slidepath)
# Gets deepzoom tile division
slide_dz = deepzoom.DeepZoomGenerator(slide,
tile_size=(size - 2),
overlap=1)
# Gets the name and number of the slide
slidename = os.path.basename(slidepath)
# Saves a preview of the slide under 'slidename.png'
slide_preview(slide, slidename, outpath)
# Asures that the chosen level is valid
if level < slide_dz.level_count:
tiles = slide_dz.level_tiles[level]
print('Level {} contains {} tiles (empty tiles included)'.format(
level,
slide_dz.level_tiles[level][0] * slide_dz.level_tiles[level][1]))
else:
print('Invalid level')
return 0
# Creates new directory - where patches will be stored
outpath = os.path.join(outpath, slidename)
try:
os.mkdir(outpath)
print("Directory", outpath, "created")
except FileExistsError:
print("Directory", outpath, "already exists")
print("Patches already extracted")
return 0
# Saves tiles if detects tissue presence higher than tissue_ratio
n = 0
print("Saving tiles image " + slidepath + "...")
for i in tqdm(range(tiles[0])):
for j in range(tiles[1]):
# Gets the tile in position (i, j)
tile = slide_dz.get_tile(level, (i, j))
image = numpy.array(tile)[..., :3]
mask = tissue.get_tissue_from_rgb(image, blacktol=10, whitetol=240)
# Saves tile in outpath only if tissue ratio is higher than threshold
if mask.sum() > tissue_ratio * tile.size[0] * tile.size[1]:
tile_path = os.path.join(
outpath,
'{}#{}-level{}-{}-{}.jpg'.format(slidename, n, level, i,
j))
tile.save(tile_path)
n = n + 1
print('Total of {} tiles with tissue ratio >{} in slide {}'.format(
n, tissue_ratio, slidepath))
print()
return n
def file_to_dask_array(path,
tile_size=1000,
overlap=0,
remove_last=True,
allow_unknown_chunksizes=False):
"""Load an image to a dask array.
Parameters
----------
path : str
The path to the image file as a string.
tile_size : int, optional
Size of chunk to be read in.
overlap : int, optional
Do not modify, overlap between neighboring tiles.
remove_last : bool, optional
Remove last tile because it has a custom size.
allow_unknown_chunksizes : bool, optional
Allow different chunk sizes, more flexible, but slowdown.
Returns
-------
arr : dask.array.Array
A Dask Array representing the contents of the image file.
Examples
--------
>>> da_img = file_to_dask_array(path)
>>> npa_img = arr.compute() # convert from dask array to numpy array
>>> pil_img = to_pil(cv2.resize(
... npa_img,
... dsize=(1440, 700),
... interpolation=cv2.INTER_CUBIC
... ))
>>> pil_img.save(test_image_name)
"""
if path.endswith('.npy'):
import dask.array as da
da.from_array(np.load(path))
else:
import openslide
img = openslide.open_slide(path)
if isinstance(img, openslide.OpenSlide):
from openslide import deepzoom
import dask.array as da
import dask.delayed
gen = deepzoom.DeepZoomGenerator(img,
tile_size=tile_size,
overlap=overlap,
limit_bounds=True)
max_level = len(gen.level_dimensions) - 1
n_tiles_x, n_tiles_y = gen.level_tiles[max_level]
@dask.delayed(pure=True)
def get_tile(level, column, row):
tile = gen.get_tile(level, (column, row)) # PIL.Image
return da.transpose(da.from_array(np.array(tile)),
axes=(1, 0, 2))
sample_tile_shape = get_tile(max_level, 0, 0).shape.compute()
rows = range(n_tiles_y - (0 if not remove_last else 1))
cols = range(n_tiles_x - (0 if not remove_last else 1))
tiles = [
da.concatenate(
[
da.from_delayed(get_tile(max_level, col, row),
sample_tile_shape, np.uint8)
for row in rows
],
allow_unknown_chunksizes=allow_unknown_chunksizes,
axis=1) for col in cols
]
arr = da.concatenate(
tiles,
allow_unknown_chunksizes=allow_unknown_chunksizes).transpose(
[1, 0, 2])
return arr
else: # img is instance of openslide.ImageSlide
import dask_image.imread
return dask_image.imread.imread(path)
