def test_patch_bk_ratio():
img_path = os.path.join(PRJ_PATH, "test/data/Images/3c32efd9.png")
img = io.imread(img_path)
bk_ratio = patch.patch_bk_ratio(img, bk_thresh=0.80)
if bk_ratio > 1 or bk_ratio < 0:
raise AssertionError("Ratio not in the range.")
def gen_wsi_feas(patch_model, img_dir, fea_dir, args):
img_list = [ele for ele in os.listdir(img_dir) if "jpg" in ele]
for ind, ele in enumerate(img_list):
img_name = os.path.splitext(ele)[0]
if ind > 0 and ind % 10 == 0:
print("processing {:03d}/{:03d}, {}".format(ind, len(img_list), img_name))
feas_list, probs_list, coor_list = [], [], []
img_path = os.path.join(img_dir, ele)
cur_img = io.imread(img_path)
# split coors and save patches
coors_arr = wsi_stride_splitting(cur_img.shape[0], cur_img.shape[1], args.patch_len, args.stride_len)
patch_list = []
for ind, coor in enumerate(coors_arr):
start_h, start_w = coor[0], coor[1]
patch_img = cur_img[start_h:start_h+args.patch_len, start_w:start_w+args.patch_len]
# image background control
if patch.patch_bk_ratio(patch_img, bk_thresh=0.864) <= 0.88:
patch_list.append(patch_img)
coor_list.append([start_h, start_w, start_h+args.patch_len, start_w+args.patch_len])
# Processing the feature extraction in batch-wise manner to avoid huge memory consumption
if len(patch_list) == 16 or ind+1 == len(coors_arr):
patch_arr = np.asarray(patch_list)
patch_dset = PatchDataset(patch_arr)
patch_loader = DataLoader(patch_dset, batch_size=16, shuffle=False, num_workers=4, drop_last=False)
with torch.no_grad():
for inputs in patch_loader:
batch_tensor = Variable(inputs.cuda())
feas, probs = extract_model_feas(patch_model, batch_tensor, args)
batch_feas = feas.cpu().data.numpy().tolist()
batch_probs = probs.cpu().data.numpy().tolist()
feas_list.extend(batch_feas)
probs_list.extend(batch_probs)
patch_list = []
all_feas = np.asarray(feas_list).astype(np.float32)
all_probs = np.asarray(probs_list).astype(np.float32)
all_coors = np.asarray(coor_list).astype(np.float32)
sorted_ind = np.argsort(all_probs[:, 0])
sorted_feas = all_feas[sorted_ind]
sorted_probs = all_probs[sorted_ind]
sorted_coors = all_coors[sorted_ind]
if len(feas_list) != len(probs_list) or len(feas_list) != len(coor_list):
print("{} feas/probs/coors not consistent.".format(img_name))
else:
patch_fea_dict = {
"feas": sorted_feas,
"probs": sorted_probs,
"coors": sorted_coors,
}
if not os.path.exists(args.fea_dir):
os.makedirs(args.fea_dir)
dd.io.save(os.path.join(args.fea_dir, img_name+".h5"), patch_fea_dict)
def gen_wsi_feas(patch_model, img_path, args):
img_name = os.path.splitext(img_path)[0]
feas_list, probs_list, coor_list = [], [], []
cur_img = io.imread(img_path)
# split coors and save patches
coors_arr = wsi_stride_splitting(cur_img.shape[0], cur_img.shape[1],
args.patch_len, args.stride_len)
patch_list = []
for ind, coor in enumerate(coors_arr):
start_h, start_w = coor[0], coor[1]
patch_img = cur_img[start_h:start_h + args.patch_len,
start_w:start_w + args.patch_len]
# image background control
if patch.patch_bk_ratio(patch_img, bk_thresh=0.864) <= 0.88:
patch_list.append(patch_img)
coor_list.append([
start_h, start_w, start_h + args.patch_len,
start_w + args.patch_len
])
# Processing the feature extraction in batch-wise manner to avoid huge memory consumption
if len(patch_list) == args.cls_batch_size or ind + 1 == len(coors_arr):
patch_arr = np.asarray(patch_list)
patch_dset = ClsPatchDataset(patch_arr)
patch_loader = DataLoader(patch_dset,
batch_size=args.cls_batch_size,
shuffle=False,
num_workers=0,
drop_last=False)
with torch.no_grad():
for inputs in patch_loader:
batch_tensor = Variable(inputs.cuda())
feas, probs = extract_model_feas(patch_model, batch_tensor,
args)
batch_feas = feas.cpu().data.numpy().tolist()
batch_probs = probs.cpu().data.numpy().tolist()
feas_list.extend(batch_feas)
probs_list.extend(batch_probs)
patch_list = []
all_feas = np.asarray(feas_list).astype(np.float32)
all_probs = np.asarray(probs_list).astype(np.float32)
sorted_ind = np.argsort(all_probs[:, 0])
feas_placeholder = np.zeros((args.wsi_patch_num, all_feas.shape[1]),
dtype=np.float32)
test_patch_num = min(len(all_feas), args.wsi_patch_num)
chosen_total_ind = sorted_ind[:test_patch_num]
feas_placeholder[:test_patch_num] = all_feas[chosen_total_ind]
chosen_coors = np.asarray(coor_list)[chosen_total_ind].tolist()
return feas_placeholder, test_patch_num, chosen_coors
def gen_patches(img_dir, patch_dir, patch_size=448):
img_list = pydaily.filesystem.find_ext_files(img_dir, "jpg")
img_list = [os.path.basename(ele) for ele in img_list]
pos_patch_dir = os.path.join(patch_dir, "1Pos")
if not os.path.exists(pos_patch_dir):
os.makedirs(pos_patch_dir)
neg_patch_dir = os.path.join(patch_dir, "0Neg")
if not os.path.exists(neg_patch_dir):
os.makedirs(neg_patch_dir)
pos_num, neg_num = 0, 0
for ind, ele in enumerate(img_list):
if ind > 0 and ind % 10 == 0:
print("processing {}/{}".format(ind, len(img_list)))
img_path = os.path.join(img_dir, ele)
mask_path = os.path.join(img_dir, os.path.splitext(ele)[0] + ".png")
cur_img = io.imread(img_path)
cur_mask = io.imread(mask_path)
# split coors and save patches
coors_arr = patch.wsi_coor_splitting(cur_img.shape[0],
cur_img.shape[1],
patch_size,
overlap_flag=True)
for coor in coors_arr:
start_h, start_w = coor[0], coor[1]
patch_img = cur_img[start_h:start_h + patch_size,
start_w:start_w + patch_size]
# image background control
if patch.patch_bk_ratio(patch_img, bk_thresh=0.864) > 0.88:
continue
# mask control
patch_mask = cur_mask[start_h:start_h + patch_size,
start_w:start_w + patch_size]
pixel_ratio = np.sum(patch_mask > 0) * 1.0 / patch_mask.size
patch_name = str(uuid.uuid4())[:8]
if pixel_ratio >= 0.05:
io.imsave(os.path.join(pos_patch_dir, patch_name + ".png"),
patch_img)
pos_num += 1
else:
if np.random.random_sample() <= 0.80 and "neg" in img_dir:
continue
io.imsave(os.path.join(neg_patch_dir, patch_name + ".png"),
patch_img)
neg_num += 1
print("There are {} pos samples and {} neg samples".format(
pos_num, neg_num))
def gen_samples(slides_dir, patch_level, patch_size, tumor_type, slide_list, dset, overlap_mode):
# prepare saving directory
patch_path = os.path.join(os.path.dirname(slides_dir), "Patches", tumor_type)
patch_img_dir = os.path.join(patch_path, dset, "imgs")
if not os.path.exists(patch_img_dir):
os.makedirs(patch_img_dir)
patch_mask_dir = os.path.join(patch_path, dset, "masks")
if not os.path.exists(patch_mask_dir):
os.makedirs(patch_mask_dir)
# processing slide one-by-one
ttl_patch = 0
slide_list.sort()
for ind, ele in enumerate(slide_list):
print("Processing {} {}/{}".format(ele, ind+1, len(slide_list)))
cur_slide_path = os.path.join(slides_dir, ele+".svs")
if os.path.exists(cur_slide_path):
cur_slide_path = os.path.join(slides_dir, ele+".svs")
# locate contours and generate batches based on tissue contours
cnts, d_factor = tl.locate_tissue_cnts(cur_slide_path, max_img_size=2048, smooth_sigma=13,
thresh_val=0.88, min_tissue_size=120000)
select_level, select_factor = tl.select_slide_level(cur_slide_path, max_size=2048)
cnts = sorted(cnts, key=lambda x: cv2.contourArea(x), reverse=True)
# scale contour to slide level 2
wsi_head = pyramid.load_wsi_head(cur_slide_path)
cnt_scale = select_factor / int(wsi_head.level_downsamples[patch_level])
tissue_arr = cv_cnt_to_np_arr(cnts[0] * cnt_scale).astype(np.int32)
# convert tissue_arr to convex if poly is not valid
tissue_poly = np_arr_to_poly(tissue_arr)
if tissue_poly.is_valid == False:
tissue_arr = poly_to_np_arr(tissue_poly.convex_hull).astype(int)
coors_arr = None
if overlap_mode == "half_overlap":
level_w, level_h = wsi_head.level_dimensions[patch_level]
coors_arr = contour.contour_patch_splitting_half_overlap(tissue_arr, level_h, level_w, patch_size, inside_ratio=0.80)
elif overlap_mode == "self_overlap":
coors_arr = contour.contour_patch_splitting_self_overlap(tissue_arr, patch_size, inside_ratio=0.80)
else:
raise NotImplementedError("unknown overlapping mode")
wsi_img = wsi_head.read_region((0, 0), patch_level, wsi_head.level_dimensions[patch_level])
wsi_img = np.asarray(wsi_img)[:,:,:3]
mask_path = os.path.join(slides_dir, "_".join([ele, tumor_type+".tif"]))
mask_img = io.imread(mask_path)
wsi_mask = (transform.resize(mask_img, wsi_img.shape[:2], order=0) * 255).astype(np.uint8) * 255
if dset == "val":
test_slides_dir = os.path.join(os.path.dirname(slides_dir), "TestSlides")
if not os.path.exists(os.path.join(test_slides_dir, cur_slide_path)):
shutil.copy(cur_slide_path, test_slides_dir)
if not os.path.exists(os.path.join(test_slides_dir, mask_path)):
shutil.copy(mask_path, test_slides_dir)
for cur_arr in coors_arr:
cur_h, cur_w = cur_arr[0], cur_arr[1]
cur_patch = wsi_img[cur_h:cur_h+patch_size, cur_w:cur_w+patch_size]
if cur_patch.shape[0] != patch_size or cur_patch.shape[1] != patch_size:
continue
cur_mask = wsi_mask[cur_h:cur_h+patch_size, cur_w:cur_w+patch_size]
# background RGB (235, 210, 235) * [0.299, 0.587, 0.114]
if patch.patch_bk_ratio(cur_patch, bk_thresh=0.864) > 0.88:
continue
if overlap_mode == "half_overlap" and tumor_type == "viable":
pixel_ratio = np.sum(cur_mask > 0) * 1.0 / cur_mask.size
if pixel_ratio < 0.05:
continue
patch_name = ele + "_" + str(uuid.uuid1())[:8]
io.imsave(os.path.join(patch_img_dir, patch_name+".jpg"), cur_patch)
io.imsave(os.path.join(patch_mask_dir, patch_name+".png"), cur_mask)
ttl_patch += 1
print("There are {} patches in total.".format(ttl_patch))