def process():
"""
Performs post processing for a list of images
"""
cfg = Config()
for data_dir in cfg.inf_data_list:
proc_dir = cfg.inf_output_dir + '/processed/'
pred_dir = cfg.inf_output_dir + '/raw/'
file_list = glob.glob(pred_dir + '*.npy')
file_list.sort() # ensure same order
if not os.path.isdir(proc_dir):
os.makedirs(proc_dir)
for filename in file_list:
start = time.time()
filename = os.path.basename(filename)
basename = filename.split('.')[0]
test_set = basename.split('_')[0]
test_set = test_set[-1]
print(pred_dir, basename, end=' ', flush=True)
##
img = cv2.imread(data_dir + basename + cfg.inf_imgs_ext)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
pred_map = np.load(pred_dir + '/%s.npy' % basename)
# get the instance level prediction
pred_inst = process_utils(pred_map, cfg.model_mode)
# ! remap label is slow - check to see whether it is needed!
pred_inst = remap_label(pred_inst, by_size=True)
overlaid_output = visualize_instances(pred_inst, img)
overlaid_output = cv2.cvtColor(overlaid_output, cv2.COLOR_BGR2RGB)
cv2.imwrite('%s/%s.png' % (proc_dir, basename), overlaid_output)
# save segmentation mask
np.save('%s/%s' % (proc_dir, basename), pred_inst)
end = time.time()
diff = str(round(end - start, 2))
print('FINISH. TIME: %s' % diff)
if not os.path.isdir(proc_dir):
os.makedirs(proc_dir)
for filename in file_list:
filename = os.path.basename(filename)
basename = filename.split('.')[0]
print(basename, norm_target, end=' ')
##
img = cv2.imread(imgs_dir + basename + cfg.inf_imgs_ext)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
pred = sio.loadmat('%s/%s.mat' % (pred_dir, basename))
pred = np.squeeze(pred['result'])
if proc_mode == 'np+xy':
pred = proc_np_xy(pred,
marker_mode=marker_mode,
energy_mode=energy_mode)
if proc_mode == 'np+dst':
pred = proc_np_dst(pred)
##
overlaid_output = visualize_instances(pred, img)
overlaid_output = cv2.cvtColor(overlaid_output, cv2.COLOR_BGR2RGB)
cv2.imwrite('%s/%s.png' % (proc_dir, basename), overlaid_output)
sio.savemat('%s/%s_predicted_map.mat' % (proc_dir, basename),
{'predicted_map': pred})
##
print('FINISH')
if cfg.model_type == 'np_hv':
pred_inst = postproc.hover.proc_np_hv(pred_inst,
marker_mode=marker_mode,
energy_mode=energy_mode, rgb=img)
elif cfg.model_type == 'np_dist':
pred_inst = postproc.hover.proc_np_dist(pred_inst)
elif cfg.model_type == 'dist':
pred_inst = postproc.dist.process(pred_inst)
else:
pred_inst = postproc.other.process(pred_inst, cfg.model_type)
# ! will be extremely slow on WSI/TMA so it's advisable to comment this out
# * remap once so that further processing faster (metrics calculation, etc.)
pred_inst = remap_label(pred_inst, by_size=True)
overlaid_output = visualize_instances(pred_inst, img)
# plt.subplot(1,2,1)
# plt.imshow(pred_inst)
# plt.subplot(1,2,2)
# plt.imshow(pred_type)
# plt.imshow()
# np.save('/home/test/GhulamMurtaza/pred_maps_class_instance/%s.npy' % (basename),pred_type)
# np.save('/home/test/GhulamMurtaza/pred_maps_class_instance/%s.npy' % (basename),pred_inst)
overlaid_output_type = visualize_instances(pred_type,img,color=[[255,0,0],[255,255,0],[0,0,225],[0,128,0],[100,122,56]])
overlaid_output_true = visualize_instances(true_mask_type,img,color=[[255,0,0],[255,255,0],[0,0,225],[0,128,0],[100,122,56]])
overlaid_output = cv2.cvtColor(overlaid_output, cv2.COLOR_BGR2RGB)
overlaid_output_type = cv2.cvtColor(overlaid_output_type,cv2.COLOR_BGR2RGB)
overlaid_output_true = cv2.cvtColor(overlaid_output_true,cv2.COLOR_BGR2RGB)
cv2.imwrite('%s/%s.png' % (proc_dir, basename), overlaid_output)
def process_image(filename):
filename = os.path.basename(filename)
basename = filename.split(".")[0]
if jobs == 1:
print(pred_dir, basename, flush=True)
##
img = cv2.imread(
os.path.join(data_dir, "{}{}".format(basename,
cfg.inf_imgs_ext)))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
pred = sio.loadmat(
os.path.join(pred_dir, "{}.mat".format(basename)))
pred = np.squeeze(pred["result"])
if hasattr(cfg, "type_classification") and cfg.type_classification:
pred_inst = pred[..., cfg.nr_types:]
pred_type = pred[..., :cfg.nr_types]
pred_inst = np.squeeze(pred_inst)
pred_type = np.argmax(pred_type, axis=-1)
else:
pred_inst = pred
if cfg.model_type == "np_hv" or cfg.model_type == "np_hv_opt":
pred_inst = postproc.hover.proc_np_hv(pred_inst,
marker_mode=marker_mode,
energy_mode=energy_mode,
rgb=img)
elif cfg.model_type == "np_dist":
pred_inst = postproc.hover.proc_np_dist(pred_inst)
# ! will be extremely slow on WSI/TMA so it's advisable to comment this out
# * remap once so that further processing faster (metrics calculation, etc.)
if cfg.remap_labels:
pred_inst = remap_label(pred_inst, by_size=True)
# for instance segmentation only
if cfg.type_classification:
#### * Get class of each instance id, stored at index id-1
pred_id_list = list(
np.unique(pred_inst))[1:] # exclude background ID
pred_inst_type = np.full(len(pred_id_list), 0, dtype=np.int32)
for idx, inst_id in enumerate(pred_id_list):
inst_type = pred_type[pred_inst == inst_id]
type_list, type_pixels = np.unique(inst_type,
return_counts=True)
type_list = list(zip(type_list, type_pixels))
type_list = sorted(type_list,
key=lambda x: x[1],
reverse=True)
inst_type = type_list[0][0]
if inst_type == 0: # ! pick the 2nd most dominant if exist
if len(type_list) > 1:
inst_type = type_list[1][0]
else:
if jobs == 1:
pass # print('[Warn] Instance has `background` type')
pred_inst_type[idx] = inst_type
pred_inst_centroid = get_inst_centroid(pred_inst)
###### ad hoc just once for pannuke predictions
# for key in ['type_map', 'inst_type']:
# pred_type[(pred_type == 5)] = 4
# pred_inst_type[(pred_inst_type == 5)] = 4
sio.savemat(
os.path.join(proc_dir, "{}.mat".format(basename)),
{
"inst_map": pred_inst,
"type_map": pred_type,
"inst_type": pred_inst_type[:, None],
"inst_centroid": pred_inst_centroid,
},
)
overlaid_output = visualize_instances(
pred_inst,
img,
((cfg.nuclei_type_dict, cfg.color_palete),
pred_inst_type[:, None]),
cfg.outline,
cfg.skip_types,
)
overlaid_output = cv2.cvtColor(overlaid_output,
cv2.COLOR_BGR2RGB)
cv2.imwrite(os.path.join(proc_dir, "{}.png".format(basename)),
overlaid_output)
with open(os.path.join(proc_dir, f"{basename}.log"),
"w") as log_file:
unique, counts = np.unique(pred_inst_type[:, None],
return_counts=True)
unique = list(unique)
if 0 in unique: # remove backround entries
counts = np.delete(counts, unique.index(0))
unique.remove(0)
print(
f"{basename} : {dict(zip([{str(v): str(k) for k, v in cfg.nuclei_type_dict.items()}[str(item)] for item in unique], counts))}",
file=log_file,
)
else:
sio.savemat(
os.path.join(proc_dir, "{}.mat".format(basename)),
{"inst_map": pred_inst},
)
overlaid_output = visualize_instances(pred_inst, img)
overlaid_output = cv2.cvtColor(overlaid_output,
cv2.COLOR_BGR2RGB)
cv2.imwrite(
os.path.join(proc_dir, "{}_uc.png".format(basename)),
overlaid_output,
)
##
if jobs == 1:
print(
f"Finished for {basename} {datetime.now().strftime('%H:%M:%S.%f')}"
)
cv2.imwrite('%s/%s.png' % (save_dir, basename), patch)
save_dir = os.path.join(output_folder, 'Labels')
if not os.path.exists(save_dir): os.mkdir(save_dir)
inst_centroids *= scale
sio.savemat(
'%s/%s.mat' % (save_dir, basename), {
'inst_map': instance_map,
'type_map': type_map,
'inst_type': inst_type[:, None],
'inst_centroid': inst_centroids,
})
save_dir = os.path.join(output_folder, 'Combined')
if not os.path.exists(save_dir): os.mkdir(save_dir)
colormap = [class_colors[x] for x in inst_type]
overlaid_output = visualize_instances(instance_map, patchcorr,
colormap)
cv2.imwrite("{}/{}.png".format(save_dir, basename), overlaid_output)
#type_map[np.logical_and(instance_map > 0, type_map == 0)] = -1
unannot = np.logical_and(instance_map != 0, type_map == 0)
patchjoin[unannot] = 255
instance_map[unannot] = 0
ann = np.dstack([instance_map, type_map])
patchjoin = np.array(patchjoin, np.int32)
print(ann.shape)
print(patchjoin.shape)
patchjoin = np.concatenate([patchjoin, ann], axis=-1)
sub_patches = xtractor.extract(
patchjoin, extract_type
) # This extracts mirrored subpatch of 540x540 from 1k patch
for idx, patch in enumerate(sub_patches):
np.save("{}/{}_{}.npy".format(save_dir, basename, idx), patch)
def process_image(filename):
filename = os.path.basename(filename)
basename = filename.split('.')[0]
if not parallel: print(pred_dir, basename, flush=True)
##
img = cv2.imread(os.path.join(data_dir, '{}{}'.format(basename, cfg.inf_imgs_ext)))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
pred = sio.loadmat(os.path.join(pred_dir, '{}.mat'.format(basename)))
pred = np.squeeze(pred['result'])
if hasattr(cfg, 'type_classification') and cfg.type_classification:
pred_inst = pred[...,cfg.nr_types:]
pred_type = pred[...,:cfg.nr_types]
pred_inst = np.squeeze(pred_inst)
pred_type = np.argmax(pred_type, axis=-1)
if cfg.model_type == 'micronet':
# dilate prediction of all type to match it with
# the instance segmentation post-proc code
kernel = np.array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], np.uint8)
canvas = np.zeros_like(pred_type, dtype=np.int32)
for type_id in range(1, cfg.nr_classes):
type_map = (pred_type == type_id).astype('uint8')
type_map = cv2.dilate(type_map, kernel, iterations=1)
canvas[type_map > 0] = type_id
else:
pred_inst = pred
if cfg.model_type == 'np_hv' or cfg.model_type == 'np_hv_opt':
pred_inst = postproc.hover.proc_np_hv(pred_inst,
marker_mode=marker_mode,
energy_mode=energy_mode, rgb=img)
elif cfg.model_type == 'np_dist':
pred_inst = postproc.hover.proc_np_dist(pred_inst)
elif cfg.model_type == 'dist':
pred_inst = postproc.dist.process(pred_inst)
else:
pred_inst = postproc.other.process(pred_inst, cfg.model_type)
# ! will be extremely slow on WSI/TMA so it's advisable to comment this out
# * remap once so that further processing faster (metrics calculation, etc.)
if (cfg.remap_labels):
pred_inst = remap_label(pred_inst, by_size=True)
# for instance segmentation only
if cfg.type_classification:
#### * Get class of each instance id, stored at index id-1
pred_id_list = list(np.unique(pred_inst))[1:] # exclude background ID
pred_inst_type = np.full(len(pred_id_list), 0, dtype=np.int32)
for idx, inst_id in enumerate(pred_id_list):
inst_type = pred_type[pred_inst == inst_id]
type_list, type_pixels = np.unique(inst_type, return_counts=True)
type_list = list(zip(type_list, type_pixels))
type_list = sorted(type_list, key=lambda x: x[1], reverse=True)
inst_type = type_list[0][0]
if inst_type == 0: # ! pick the 2nd most dominant if exist
if len(type_list) > 1:
inst_type = type_list[1][0]
else:
if not parallel:
pass # print('[Warn] Instance has `background` type')
pred_inst_type[idx] = inst_type
pred_inst_centroid = get_inst_centroid(pred_inst)
###### ad hoc just once for pannuke predictions
# for key in ['type_map', 'inst_type']:
# pred_type[(pred_type == 5)] = 4
# pred_inst_type[(pred_inst_type == 5)] = 4
sio.savemat(os.path.join(proc_dir, '{}.mat'.format(basename)),
{'inst_map' : pred_inst,
'type_map' : pred_type,
'inst_type' : pred_inst_type[:, None],
'inst_centroid' : pred_inst_centroid,
})
overlaid_output = visualize_instances(pred_inst, img, ((cfg.nuclei_type_dict, cfg.color_palete), pred_inst_type[:, None]), cfg.outline, cfg.skip_types)
overlaid_output = cv2.cvtColor(overlaid_output, cv2.COLOR_BGR2RGB)
cv2.imwrite(os.path.join(proc_dir, '{}.png'.format(basename)), overlaid_output)
with open(os.path.join(proc_dir, f'{basename}.log'), 'w') as log_file:
unique, counts = np.unique(pred_inst_type[:, None], return_counts=True)
unique = list(unique)
if 0 in unique: # remove backround entries
counts = np.delete(counts, unique.index(0))
unique.remove(0)
print(f'{basename} : {dict(zip([{str(v): str(k) for k, v in cfg.nuclei_type_dict.items()}[str(item)] for item in unique], counts))}', file = log_file)
else:
sio.savemat(os.path.join(proc_dir, '{}.mat'.format(basename)),
{'inst_map' : pred_inst})
overlaid_output = visualize_instances(pred_inst, img)
overlaid_output = cv2.cvtColor(overlaid_output, cv2.COLOR_BGR2RGB)
cv2.imwrite(os.path.join(proc_dir, '{}_uc.png'.format(basename)), overlaid_output)
##
if not parallel: print(f"Finished for {basename} {datetime.now().strftime('%H:%M:%S.%f')}")
