def visualize_prediction_came(feature_name, pred_results, M, labels, train_len=0, output_prefix=""):
data_counter = 0
validation_offset = 0
for m in M:
metadata = m
scalars = np.abs(pred_results * 100. / np.max(pred_results))
image_root = "/datagrid/Medical/microscopy/CAMELYON16/training"
image_name = os.path.basename(metadata[0])
patch_size = 1024 # int(os.path.basename(metadata[0]).split('_')[3][1:])
image_name_parts = image_name.split('_')
if 'tumor' in metadata[0]:
image_root += "/tumor/t"
image_path = image_root + "{}_{}.tif".format(image_name_parts[1][1:], image_name_parts[2])
elif 'normal' in metadata[0]:
image_root += "/normal/N"
image_path = image_root + "{}_{}.tif".format(image_name_parts[1][1:], image_name_parts[2])
continue
else:
image_root = "/datagrid/Medical/microscopy/CAMELYON16/testing/"
image_path = image_root + "Test_{}.tif".format(image_name_parts[2])
si = CamelyonSlideImage(image_path, None)
print(" [VisOutput] Loading image {}".format(image_path))
if not os.path.exists(image_path):
print (" !! FAILED, IMAGE NOT FOUND !!")
break
patchvis = si.get_patch_visualization(6, metadata[2], patch_size,
scalars=scalars[
validation_offset + m[3]:validation_offset + m[3] + len(
m[2])],
line_thickness=1,
show=False, filled=True)
# if data_counter < train_len:
# cv2.putText(patchvis, 'test', (40, 40), cv2.FONT_HERSHEY_PLAIN, 1, (255, 0, 0), 2, cv2.LINE_AA)
# else:
# cv2.putText(patchvis, 'vali', (40, 40), cv2.FONT_HERSHEY_PLAIN, 1, (0, 0, 255), 2, cv2.LINE_AA)
cv2.imwrite("{}_{}_{}".format(output_prefix, image_name, feature_name) + ".png", patchvis)
data_counter += 1
parser.add_argument("-o", "--output_prefix", type=str, help='Output image prefix', default=None, required=True)
parser.add_argument("-m", "--mask", type=str, help='Mask image, tissue class with max value within',
default=None, required=True)
parser.add_argument("-ml", "--mask-level", type=int, help='Extraction level of the mask image',
default=6, required=True)
parser.add_argument("-t", "--tumor-annot", type=str,
help='Tumor annotation file (default: searching for standard locations/extensions)',
default=None, required=False)
opt_group = parser.add_argument_group("Optional arguments")
opt_group.add_argument("-l", "--extract-level", type=int, help='Level for extracting patches',
default=3, required=False)
cmdline_args = parser.parse_args()
si = CamelyonSlideImage(image_path=cmdline_args.image,
tissue_mask_path=cmdline_args.mask, tumor_annotation_file=cmdline_args.tumor_annot)
tissue_mask_img = cv2.imread(cmdline_args.mask, cv2.IMREAD_GRAYSCALE)
tissue_mask_img -= np.min(tissue_mask_img)
im2, contours, hierarchy = cv2.findContours(tissue_mask_img,
mode=cv2.RETR_CCOMP,
method=cv2.CHAIN_APPROX_NONE)
for ci, c in enumerate(contours):
minx = im2.size
miny = im2.size
maxx = 0
maxy = 0
opt_group.add_argument("-md",
"--modulation-matrix",
type=str,
help='Modulation matrix for separation optimization'
' (add option \'randsep\' to the list \'-e\'',
default=None,
required=False)
opt_group.add_argument("--preview",
help='Flag whether to show intermediate results',
action='store_true')
cmdline_args = parser.parse_args()
si = None
if '.ndpi' not in cmdline_args.image:
si = CamelyonSlideImage(image_path=cmdline_args.image,
tissue_mask_path=cmdline_args.mask)
is_camelyon = True
rgb_from_her = np.array([[0.7595018, 0.51920101, 0.38165572],
[0.4895436, 0.74380669, 0.50788103],
[0.53518641, 0.76756465, 0.35352657]])
modulation_matrix = None
if cmdline_args.convolution_matrix is not None:
rgb_from_her = eval(cmdline_args.convolution_matrix)
print('Retrieved color-mixing matrix: \n ' + str(rgb_from_her))
her_from_rgb = np.linalg.inv(rgb_from_her)
if cmdline_args.modulation_matrix is not None:
modulation_matrix = eval(cmdline_args.modulation_matrix)
print('Retrieved modulation matrix: \n ' + str(modulation_matrix))
def visualize_prediction(algorithms,
pred_results,
M,
labels,
train_len=0,
output_prefix=""):
visualize_prediction.counter += 1
data_counter = 0
validation_offset = 0
for m in M:
metadata = m
for algorithm in algorithms:
alg_name = algorithm[2]
# mean_preds = np.sum(np.array(pred_results[alg_name])-np.array(labels), axis=0)
mean_preds = np.sum(np.array(pred_results[alg_name]), axis=0)
max_preds = np.max(np.array(pred_results[alg_name]), axis=0)
min_preds = np.min(np.array(pred_results[alg_name]), axis=0)
sd_preds = np.sum(np.array(pred_results[alg_name]), axis=0)
# sd_preds = np.std(np.array(pred_results[alg_name])-np.array(labels), axis=0)
scalars = np.abs(mean_preds * 100. / np.max(mean_preds))
border_scalars = np.zeros_like(scalars)
for li, label in enumerate(labels):
if label > 0:
border_scalars[li] = 100 * min_preds[li]
else:
border_scalars[li] = 100 * max_preds[li]
image_root = "/datagrid/Medical/microscopy/CAMELYON16/training"
image_name = os.path.basename(metadata[0])
image_name_parts = image_name.split('_')
patch_size = int(image_name_parts[5][1:])
if 'tumor' in metadata[0]:
image_root += "/tumor/"
image_path = image_root + "tumor_{}.tif".format(
image_name_parts[2])
elif 'normal' in metadata[0]:
image_root += "/normal/"
image_path = image_root + "Normal_{}.tif".format(
image_name_parts[2])
si = CamelyonSlideImage(image_path, None)
print(" [VisOutput] Loading image {}".format(image_path))
if not os.path.exists(image_path):
print(" !! FAILED, IMAGE NOT FOUND !!")
break
patchvis = si.get_patch_visualization(
6,
metadata[2],
patch_size,
scalars=scalars[validation_offset + m[3]:validation_offset +
m[3] + len(m[2])],
border_scalars=border_scalars[validation_offset +
m[3]:validation_offset + m[3] +
len(m[2])],
line_thickness=1,
show=False,
filled=True)
if data_counter < train_len:
cv2.putText(patchvis, 'test', (40, 40), cv2.FONT_HERSHEY_PLAIN,
1, (255, 0, 0), 2, cv2.LINE_AA)
else:
cv2.putText(patchvis, 'vali', (40, 40), cv2.FONT_HERSHEY_PLAIN,
1, (0, 0, 255), 2, cv2.LINE_AA)
data_counter += 1
if data_counter == train_len:
validation_offset = m[3] + len(m[2])
cv2.imwrite(
"/local/temporary/clf_vis/{}_l0_vis_".format(output_prefix) +
image_name + "_" + alg_name +
"features_{:d}.png".format(visualize_prediction.counter),
patchvis)
if len(patches) < 1:
print("Nothing to do for input directory " + input_file)
quit()
flush_event_id = 0
flush_time = 600
image_features = []
n = len(patches)
i = 1
avg_time = 0
cum_time = 0
dsigma = np.array([0.25, 0.125, 0.0625])
si = CamelyonSlideImage(patches[0][3], None)
is_camelyon = True
area_low = 100
if patches[0][3].find('.ndpi') > -1:
si = NDPISlideImage(patches[0][3], None)
is_camelyon = False
area_low = 50
show_intermediate = False
if 'show' in options:
show_intermediate = True
si._visualize = True
st_elem = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
st_elem_r1 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (1, 1))
if lsplit[0] == '{}.tif'.format(
patches[0][0]) and lsplit[1] is not 'None':
rgb_from_her = eval(lsplit[1])
break
if rgb_from_her is not None:
her_from_rgb = np.linalg.inv(rgb_from_her)
show_intermediate = False
level_prefix = 'L0'
area_low = 100
st_elem = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
st_elem_r1 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (1, 1))
si = CamelyonSlideImage(fpath, None)
for patch_descriptor in patches:
p_location = patch_descriptor[1]
p_size = patch_descriptor[2]
label = patch_descriptor[3]
structural_features = {}
textural_features = {}
orig_patch = si.load_patch(p_location, p_size, base_extract_level)
ori_patch = cv2.cvtColor(np.copy(orig_patch), cv2.COLOR_RGBA2RGB)
ori_patch_mf = cv2.medianBlur(ori_patch, 3)
if p_size[0] // pow(2, base_extract_level) < 256:
orig_patch_in = ori_patch_mf
"--csv",
type=str,
help='CSV file for saving the complete patch metadata',
default=None,
required=False)
opt_group.add_argument(
"--preview",
type=str,
help='File path for storing the preview image with annotated patches',
default=None,
required=False)
cmdline_args = parser.parse_args()
si = CamelyonSlideImage(image_path=cmdline_args.image,
tissue_mask_path=cmdline_args.mask,
tumor_annotation_file=cmdline_args.tumor_annot)
if '.ndpi' in cmdline_args.image:
si = NDPISlideImage(image_path=cmdline_args.image,
tissue_mask_path=cmdline_args.mask,
tumor_annotation_file=cmdline_args.tumor_annot)
p_shift = cmdline_args.patch_stride
if p_shift < 0:
p_shift = cmdline_args.patch_size
p = si.get_annotated_patches(
extract_level=cmdline_args.extract_level,
min_coverage_extraction=cmdline_args.min_coverage,
min_tumor_coverage=0.6,
p_size=cmdline_args.patch_size,
image_features = []
n = len(patches)
i = 1
avg_time = 0
cum_time = 0
dsigma = np.array([0.25, 0.125, 0.0625])
show_intermediate = False
if 'show' in options:
show_intermediate = True
if patch_extract_level > 1:
dsigma *= 2 * patch_extract_level
si = CamelyonSlideImage(patches[0][3], None)
is_camelyon = True
area_low = 100
if patches[0][3].find('.ndpi') > -1:
si = NDPISlideImage(patches[0][3], None)
is_camelyon = False
area_low = 70
for patch_descriptor in patches:
location = patch_descriptor[0]
p_size = patch_descriptor[1]
label = patch_descriptor[2]
if 'doublesize' in options:
p_size = (2 * p_size[0], 2 * p_size[1])
visited_no_patches = 0
if 0 < max_no_patches < len(patches_in['NO']):
random.shuffle(patches_in['NO'])
patches += patches_in['NO'][:max_no_patches]
else:
patches += patches_in['NO']
if len(patches) < 1:
print("Nothing to do for input directory " + input_file)
quit()
n = len(patches)
i = 1
si = CamelyonSlideImage(patches[0][3], None)
is_camelyon = True
area_low = 100
if patches[0][3].find('.ndpi') > -1:
si = NDPISlideImage(patches[0][3], None)
is_camelyon = False
area_low = 50
blur_score = []
bckg_score = []
patch_vis_meta = []
tot_area = 1. / ((patches[0][1])[0] * (patches[0][1])[1])
for patch_descriptor in patches:
p_location = patch_descriptor[0]