def shapes_to_segs(
drawn_shapes_data, image_display_top_figure, image_display_side_figure,
):
masks = np.zeros_like(img)
for j, (graph_figure, (hscale, wscale)) in enumerate(
zip([image_display_top_figure, image_display_side_figure], hwscales)
):
fig = go.Figure(**graph_figure)
# we use the width and the height of the first layout image (this will be
# one of the images of the brain) to get the bounding box of the SVG that we
# want to rasterize
width, height = [fig.layout.images[0][sz] for sz in ["sizex", "sizey"]]
for i in range(seg_img.shape[j]):
shape_args = [
dict(width=width, height=height, shape=s)
for s in drawn_shapes_data[j][i]
]
if len(shape_args) > 0:
mask = shape_utils.shapes_to_mask(
shape_args,
# we only have one label class, so the mask is given value 1
1,
)
# TODO: Maybe there's a more elegant way to downsample the mask?
np.moveaxis(masks, 0, j)[i, :, :] = mask[::hscale, ::wscale]
found_segs_tensor = np.zeros_like(img)
if DEBUG_MASK:
found_segs_tensor[masks == 1] = 1
else:
# find labels beneath the mask
labels = set(seg[1 == masks])
# for each label found, select all of the segment with that label
for l in labels:
found_segs_tensor[seg == l] = 1
return found_segs_tensor
def compute_segmentations(
shapes,
img_path="assets/segmentation_img.jpg",
features=None,
shape_layers=None,
label_to_colors_args={},
):
# load original image
img = img_to_ubyte_array(img_path)
# convert shapes to mask
shape_args = [{
"width": img.shape[1],
"height": img.shape[0],
"shape": shape
} for shape in shapes]
if (shape_layers is None) or (len(shape_layers) != len(shapes)):
shape_layers = [(n + 1) for n, _ in enumerate(shapes)]
mask = shape_utils.shapes_to_mask(shape_args, shape_layers)
# do segmentation and return this
t1 = time()
clf = RandomForestClassifier(n_estimators=50,
n_jobs=-1,
max_depth=8,
max_samples=0.05)
seg, clf = fit_segmenter(mask, features, clf)
t2 = time()
print(t2 - t1)
color_seg = label_to_colors(seg, **label_to_colors_args)
# color_seg is a 3d tensor representing a colored image whereas seg is a
# matrix whose entries represent the classes
return (color_seg, seg, clf)
def compute_segmentations(
shapes,
img_path="assets/segmentation_img.jpg",
segmenter_args={},
shape_layers=None,
label_to_colors_args={},
):
# load original image
img = img_to_ubyte_array(img_path)
# convert shapes to mask
shape_args = [{
"width": img.shape[1],
"height": img.shape[0],
"shape": shape
} for shape in shapes]
if (shape_layers is None) or (len(shape_layers) != len(shapes)):
shape_layers = [(n + 1) for n, _ in enumerate(shapes)]
mask = shape_utils.shapes_to_mask(shape_args, shape_layers)
# do segmentation and return this
seg, clf = trainable_segmentation(img, mask, **segmenter_args)
color_seg = label_to_colors(seg, **label_to_colors_args)
# color_seg is a 3d tensor representing a colored image whereas seg is a
# matrix whose entries represent the classes
return (color_seg, seg, clf)
def draw_shapes_react(
drawn_shapes_data, image_display_top_figure, image_display_side_figure
):
if any(
[
e is None
for e in [
drawn_shapes_data,
image_display_top_figure,
image_display_side_figure,
]
]
):
return dash.no_update
masks = np.zeros_like(img)
for j, (graph_figure, (hscale, wscale)) in enumerate(
zip([image_display_top_figure, image_display_side_figure], hwscales)
):
fig = go.Figure(**graph_figure)
# we use the width and the height of the first layout image (this will be
# one of the images of the brain) to get the bounding box of the SVG that we
# want to rasterize
width, height = [fig.layout.images[0][sz] for sz in ["sizex", "sizey"]]
for i in range(seg_img.shape[j]):
shape_args = [
dict(width=width, height=height, shape=s)
for s in drawn_shapes_data[j][i]
]
if len(shape_args) > 0:
mask = shape_utils.shapes_to_mask(
shape_args,
# we only have one label class, so the mask is given value 1
1,
)
# TODO: Maybe there's a more elegant way to downsample the mask?
np.moveaxis(masks, 0, j)[i, :, :] = mask[::hscale, ::wscale]
found_segs_tensor = np.zeros_like(img)
if DEBUG_MASK:
found_segs_tensor[masks == 1] = 1
else:
# find labels beneath the mask
labels = set(seg[1 == masks])
# for each label found, select all of the segment with that label
for l in labels:
found_segs_tensor[seg == l] = 1
# convert to a colored image
fst_colored = image_utils.label_to_colors(
found_segs_tensor,
colormap=["#000000", "#8A2BE2"],
alpha=[0, 128],
color_class_offset=0,
)
fstc_slices = [
[
array_to_data_url(np.moveaxis(fst_colored, 0, j)[i])
for i in range(np.moveaxis(fst_colored, 0, j).shape[0])
]
for j in range(NUM_DIMS_DISPLAYED)
]
return fstc_slices