def cut_cells(self, tifname, new_dict, save_path):
"""
:param tifname: full path name of .tif/.kfb file
:param new_dict: {x_y: [[class_i, det, class_i, det, (x,y,w,h)],]}
:param save_path: image saving path (note: image is saved under class_i)
:output format: save_path/diagnosis/tifbasename/class_i/tifname_x_y_w_h.jpg (note: x, y here is relative to wsi)
"""
try:
slide = openslide.OpenSlide(tifname)
except:
slide = TSlide(tifname)
basename = os.path.splitext(os.path.basename(tifname))[0]
for x_y, boxes in new_dict.items():
for box in boxes:
# image naming: tifname_x_y_w_h.jpg
x = int(x_y.split('_')[0]) + int(box[4][0])
y = int(x_y.split('_')[1]) + int(box[4][1])
w = int(box[4][2])
h = int(box[4][3])
image_name = "{}_x{}_y{}_w{}_h{}.jpg".format(
basename, x, y, w, h)
save_path_i = os.path.join(save_path, str(box[2]))
os.makedirs(save_path_i, exist_ok=True)
image_fullname = os.path.join(save_path_i, image_name)
slide.read_region((x, y), 0,
(w, h)).convert("RGB").save(image_fullname)
slide.close()
def cut_cells(filename, labels_csv, save_path, factor, N):
labels = read_labels_csv(labels_csv)
marked_boxes = read_labels_xml(os.path.splitext(filename)[0] + ".xml")
try:
slide = openslide.OpenSlide(filename)
except:
slide = TSlide(filename)
basename = os.path.splitext(os.path.basename(filename))[0]
parent_d = os.path.basename(os.path.dirname(filename))
save_path = os.path.join(save_path, parent_d, basename)
for box in labels:
x, y, w, h, p, label = box
marked_class_i = is_overlapped(marked_boxes, box, factor)
if marked_class_i:
image_name = "1-p{:.4f}_markedAs_{}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
p, marked_class_i, basename, x, y, w, h, N)
else:
image_name = "1-p{:.4f}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
p, basename, x, y, w, h, N)
save_path_i = os.path.join(save_path, label)
os.makedirs(save_path_i, exist_ok=True)
image_fullname = os.path.join(save_path_i, image_name)
x_N, y_N = int(x + (1 - N) * w / 2), int(y + (1 - N) * h / 2)
w_N, h_N = int(N * w), int(N * h)
slide.read_region((x_N, y_N), 0,
(w_N, h_N)).convert("RGB").save(image_fullname)
slide.close()
def worker(image, start_x, start_y, height, patch_width, patch_height, delta,
patch_save_path):
"""
按行切图，并保存为图像文件，每个 worker 只负责处理一行
:param image: 图像文件地址
:param start_x: 切图起点坐标-x
:param start_y: 切图起点坐标-y
:param height: 切图最大高度
:param patch_width: 切图尺寸-宽
:param patch_height: 切图尺寸-高
:param delta: 切图步长
:param patch_save_path:
:return: 切图数量
"""
# 结果队列
queue = []
try:
slide = openslide.OpenSlide(image)
except:
slide = TSlide(image)
# # 获取图像宽，高
# width, height = slide.dimensions
try:
while start_y < height:
# 读取patch块
patch = slide.read_region((start_x, start_y), 0,
(patch_width, patch_height))
# 图像格式转换
patch = cv2.cvtColor(np.asarray(patch), cv2.COLOR_RGBA2BGR)
# 过滤
patch_gray = cv2.cvtColor(patch, cv2.COLOR_BGR2GRAY)
if cv2.Laplacian(patch_gray, cv2.CV_64F).var() > cfg.slice.THRESH:
# 生成文件路径
save_path = os.path.join(patch_save_path,
"%s_%s.jpg" % (start_x, start_y))
# 文件写入
cv2.imwrite(save_path, patch)
cv2.imwrite(
'/tmp/metadata/cells/1/%s_%s.jpg' % (start_x, start_y),
patch)
queue.append(save_path)
start_y += delta
return queue
except:
raise
def gen_np_array_mem(self,
results,
classes=cfg.darknet.classes,
det=cfg.xception.det1,
size=cfg.xception.size):
"""
:param classes: [class_i,]
:param det: the threshold of det to use certain box or not, from darknet prediction
:param size: image size to cut, default to 299, which is used in Xception/inception
:param results: dict generated after running darknet predict: {x_y: [(class_i, det, (x,y,w,h)),]}
:return:
numpy array: numpy array of each cell, in order
cell_index: {index: [x_y, [class_i, det, (x,y,w,h)]]},
index is index in numpy array,
x_y is jpg image name, it represents cell source,
box = [class_i, det, (x,y,w,h)] is cell info from darknet
"""
def resize_img(img, size):
# pad zero with short side
img_croped = img.crop(
(-((size - img.size[0]) / 2), -((size - img.size[1]) / 2),
img.size[0] + (size - img.size[0]) / 2,
img.size[1] + (size - img.size[1]) / 2))
# now, yolo output is square, only need resize
img_resized = img_croped.resize((size, size))
return img_resized
try:
slide = openslide.OpenSlide(self.input_file)
except:
slide = TSlide(self.input_file)
cell_list = []
cell_index = {}
index = 0
for x_y, boxes in results.items():
for box in boxes:
if box[0] in classes and box[1] > det:
x = int(x_y.split('_')[0]) + int(box[2][0])
y = int(x_y.split('_')[1]) + int(box[2][1])
w = int(box[2][2])
h = int(box[2][3])
cell = slide.read_region((x, y), 0, (w, h)).convert("RGB")
cell_list.append(np.array(resize_img(cell, size)))
cell_index[index] = [x_y, list(box)]
index += 1
slide.close()
# return np.asarray(cell_list), cell_index
return cell_list, cell_index
def worker(tiff_path, keys, points_dict, save_path, N):
basename = os.path.splitext(os.path.basename(tiff_path))[0].replace(
" ", "-")
try:
slide = openslide.OpenSlide(tiff_path)
except:
slide = TSlide(tiff_path)
cell_count = 0
for x_y in keys:
boxes = points_dict[x_y]
for box in boxes:
x0, y0 = x_y.split('_')
x = int(x0) + int(box[4][0])
y = int(y0) + int(box[4][1])
w = int(box[4][2])
h = int(box[4][3])
# make save dir
cell_save_dir = os.path.join(save_path, box[2])
os.makedirs(cell_save_dir, exist_ok=True)
image_name = "1-p{:.4f}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
1 - box[3], basename, x, y, w, h, N)
cell_save_path = os.path.join(cell_save_dir, image_name)
slide.read_region(
(int(x + (1 - N) * w / 2), int(y + (1 - N) * h / 2)), 0,
(int(N * w), int(N * h))).convert("RGB").save(cell_save_path)
cell_count += 1
slide.close()
return cell_count
def worker_in_memory(image, start_x, start_y, height, patch_width,
patch_height, delta):
"""
按行切图，并保存为图像文件，每个 worker 只负责处理一行
:param image: 图像文件地址
:param start_x: 切图起点坐标-x
:param start_y: 切图起点坐标-y
:param height: 切图最大高度
:param patch_width: 切图尺寸-宽
:param patch_height: 切图尺寸-高
:param delta: 切图步长
:return: 切图数量
"""
# 结果队列
queue = {}
try:
slide = openslide.OpenSlide(image)
except:
slide = TSlide(image)
try:
while start_y < height:
# 读取patch块
patch = slide.read_region((start_x, start_y), 0,
(patch_width, patch_height))
# 图像格式转换
patch = cv2.cvtColor(np.asarray(patch), cv2.COLOR_RGBA2BGR)
key = '%s_%s' % (start_x, start_y)
queue[key] = patch
start_y += delta
return queue
except:
raise
def cut_cells_p_marked_(self,
tifname,
new_dict,
save_path,
factor=0.3,
N=4):
def get_labels(xmlname):
if not os.path.isfile(xmlname):
return []
classes = {
"#aa0000": "HSIL",
"#aa007f": "ASCH",
"#005500": "LSIL",
"#00557f": "ASCUS",
"#0055ff": "SCC",
"#aa55ff": "EC",
"#ff5500": "AGC",
"#00aa00": "FUNGI",
"#00aa7f": "TRI",
"#00aaff": "CC",
"#55aa00": "ACTINO",
"#55aa7f": "VIRUS",
"#ffffff": "NORMAL",
"#000000": "MC",
"#aa00ff": "SC",
"#ff0000": "RC",
"#aa5500": "GEC"
}
DOMTree = xml.dom.minidom.parse(xmlname)
collection = DOMTree.documentElement
annotations = collection.getElementsByTagName("Annotation")
marked_boxes = []
for annotation in annotations:
colorCode = annotation.getAttribute("Color")
if not colorCode in classes:
continue
marked_box = [classes[colorCode], []]
coordinates = annotation.getElementsByTagName("Coordinate")
marked_box[1] = [(float(coordinate.getAttribute('X')),
float(coordinate.getAttribute('Y')))
for coordinate in coordinates]
marked_boxes.append(marked_box)
return marked_boxes
def is_overlapped(marked_boxes, predicted_box, factor):
for marked_box in marked_boxes:
marked_box_obj = geometry.Polygon(marked_box[1])
predicted_box_obj = geometry.box(
predicted_box[0], predicted_box[1],
predicted_box[0] + predicted_box[2],
predicted_box[1] + predicted_box[3])
if marked_box_obj.intersection(predicted_box_obj).area / (
marked_box_obj.area + predicted_box_obj.area -
marked_box_obj.intersection(predicted_box_obj).area
) >= factor:
return marked_box[0]
return ""
tiff_dict = get_tiff_dict()
if tifname not in tiff_dict:
raise Exception("XCEPTION POSTPROCESS %s NOT FOUND" % tifname)
try:
slide = openslide.OpenSlide(tiff_dict[tifname])
except:
slide = TSlide(tiff_dict[tifname])
basename = os.path.splitext(os.path.basename(tifname))[0]
parent_d = os.path.basename(os.path.dirname(tifname))
save_path = os.path.join(save_path, parent_d, basename)
marked_boxes = get_labels(os.path.splitext(tifname)[0] + ".xml")
for x_y, boxes in new_dict.items():
for box in boxes:
# image naming: tifname_x_y_w_h_p.jpg
_, x, y = re.findall(pattern, x_y)[0]
x = int(x) + int(box[4][0])
y = int(y) + int(box[4][1])
w = int(box[4][2])
h = int(box[4][3])
marked_class_i = is_overlapped(marked_boxes, (x, y, w, h),
factor)
if marked_class_i:
image_name = "1-p{:.4f}_markedAs_{}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
1 - box[3], marked_class_i, basename, x, y, w, h, N)
save_path_i = os.path.join(save_path, box[2], "marked")
else:
image_name = "1-p{:.4f}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
1 - box[3], basename, x, y, w, h, N)
save_path_i = os.path.join(save_path, box[2])
os.makedirs(save_path_i, exist_ok=True)
image_fullname = os.path.join(save_path_i, image_name)
slide.read_region(
(int(x + (1 - N) * w / 2), int(y + (1 - N) * h / 2)), 0,
(int(N * w), int(
N * h))).convert("RGB").save(image_fullname)
slide.close()
def cut_cells_p_marked(self,
tifname,
new_dict,
save_path,
factor=0.3,
N=4):
"""
:param tifname: full path name of .tif/.kfb file
:param new_dict: {x_y: [[class_i, det, class_i, det, (x,y,w,h)],]}
:param save_path: image saving path (note: image is saved under class_i)
:param factor: overlapping threshold, added marked info to image filename if overlapped
:output format: save_path/diagnosis/tifbasename/class_i/tifname_x_y_w_h.jpg (note: x, y here is relative to wsi)
(note: x, y here is relative to wsi.
p is the value of the second det.
image size is twice the annotation box.
check if the cell is marked, add marked if so.)
"""
# https://www.lfd.uci.edu/~gohlke/pythonlibs/#shapely
def get_labels(xmlname):
"""collect labeled boxes from asap xml
:param xmlname: full path name of .xml file, got from .tif/.kfb file
:output format: [[class_i, [(xi,yi),]],]
"""
if not os.path.isfile(xmlname):
return []
classes = {
"#aa0000": "HSIL",
"#aa007f": "ASCH",
"#005500": "LSIL",
"#00557f": "ASCUS",
"#0055ff": "SCC",
"#aa557f": "ADC",
"#aa55ff": "EC",
"#ff5500": "AGC1",
"#ff557f": "AGC2",
"#ff55ff": "AGC3",
"#00aa00": "FUNGI",
"#00aa7f": "TRI",
"#00aaff": "CC",
"#55aa00": "ACTINO",
"#55aa7f": "VIRUS",
"#ffffff": "NORMAL",
"#000000": "MC",
"#aa00ff": "SC",
"#ff0000": "RC",
"#aa5500": "GEC"
}
DOMTree = xml.dom.minidom.parse(xmlname)
collection = DOMTree.documentElement
annotations = collection.getElementsByTagName("Annotation")
marked_boxes = []
for annotation in annotations:
colorCode = annotation.getAttribute("Color")
if not colorCode in classes:
continue
marked_box = [classes[colorCode], []]
coordinates = annotation.getElementsByTagName("Coordinate")
marked_box[1] = [(float(coordinate.getAttribute('X')),
float(coordinate.getAttribute('Y')))
for coordinate in coordinates]
marked_boxes.append(marked_box)
return marked_boxes
def is_overlapped(marked_boxes, predicted_box, factor):
"""check if predicted box is marked already
:param marked_boxes: [[class_i, [(xi,yi),]],]
:param box: (x, y, w, h)
:param factor: overlapping threshold, added marked info to image filename if overlapped
"""
for marked_box in marked_boxes:
marked_box_obj = geometry.Polygon(marked_box[1])
predicted_box_obj = geometry.box(
predicted_box[0], predicted_box[1],
predicted_box[0] + predicted_box[2],
predicted_box[1] + predicted_box[3])
if marked_box_obj.intersection(predicted_box_obj).area / (
marked_box_obj.area + predicted_box_obj.area -
marked_box_obj.intersection(predicted_box_obj).area
) >= factor:
return marked_box[0]
return ""
try:
slide = openslide.OpenSlide(tifname)
except:
slide = TSlide(tifname)
basename = os.path.splitext(os.path.basename(tifname))[0]
parent_d = os.path.basename(os.path.dirname(tifname))
save_path = os.path.join(save_path, parent_d, basename)
marked_boxes = get_labels(os.path.splitext(tifname)[0] + ".xml")
for x_y, boxes in new_dict.items():
for box in boxes:
# image naming: tifname_x_y_w_h_p.jpg
x = int(x_y.split('_')[0]) + int(box[4][0])
y = int(x_y.split('_')[1]) + int(box[4][1])
w = int(box[4][2])
h = int(box[4][3])
marked_class_i = is_overlapped(marked_boxes, (x, y, w, h),
factor)
if marked_class_i:
image_name = "1-p{:.4f}_markedAs_{}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
1 - box[3], marked_class_i, basename, x, y, w, h, N)
save_path_i = os.path.join(save_path, box[2], "marked")
else:
image_name = "1-p{:.4f}_{}_x{}_y{}_w{}_h{}_{}x.jpg".format(
1 - box[3], basename, x, y, w, h, N)
save_path_i = os.path.join(save_path, box[2])
os.makedirs(save_path_i, exist_ok=True)
image_fullname = os.path.join(save_path_i, image_name)
slide.read_region(
(int(x + (1 - N) * w / 2), int(y + (1 - N) * h / 2)), 0,
(int(N * w), int(
N * h))).convert("RGB").save(image_fullname)
slide.close()
def run(self):
print("Initial DARKNET and XCEPTION model ...")
total = len(self.tiff_lst)
for index, tiff in enumerate(self.tiff_lst):
# 获取大图文件名，不带后缀
tiff_basename, _ = os.path.splitext(os.path.basename(tiff))
tiff_basename = tiff_basename.replace(" ", "-")
print('Process %s / %s %s ...' % (index + 1, total, tiff_basename))
# 切片文件存储路径
slice_save_path = os.path.join(self.slice_dir_path, tiff_basename)
t0 = datetime.datetime.now()
# 如果路径下切图文件不存在，执行切图
if not os.path.exists(slice_save_path):
# 执行切图
ImageSlice(tiff, self.slice_dir_path).get_slices()
# 获取切图文件路径
tif_images = FilesScanner(slice_save_path, ['.jpg']).get_files()
t1 = datetime.datetime.now()
print('TIFF SLICE COST: %s' % (t1 - t0))
tasks = []
# 创建切图进程池
executor = ProcessPoolExecutor(max_workers=GPU_NUM)
if len(tif_images) < cfg.darknet.min_job_length:
tasks.append(executor.submit(yolo_predict, '0', tif_images))
else:
# 任务切分
n = int((len(tif_images) / float(GPU_NUM)) + 0.5)
patches = [tif_images[i: i + n] for i in range(0, len(tif_images), n)]
for gpu_index, patch in enumerate(patches):
tasks.append(executor.submit(yolo_predict, str(gpu_index), patch))
seg_results = {}
for future in as_completed(tasks):
result = future.result()
seg_results.update(result)
# 关闭进程池
executor.shutdown(wait=True)
try:
slide = openslide.OpenSlide(tiff)
except:
slide = TSlide(tiff)
keys = list(seg_results.keys())
for key in keys:
lst = seg_results[key]
x0, y0 = key.split('_')
x0, y0 = int(x0), int(y0)
for item in lst:
label, accuracy, (x, y, w, h) = item
accuracy, x, y, w, h = float(accuracy), int(x), int(y), int(w), int(h)
x, y = x0 + x, y0 + y
save_path = os.path.join(self.cells_path, tiff_basename, label)
if not os.path.exists(save_path):
os.makedirs(save_path)
image_name = "1-p{:.4f}_{}_x{}_y{}_w{}_h{}.jpg".format(1 - accuracy, tiff_basename, x, y, w, h)
slide.read_region((x, y), 0, (w, h)).convert("RGB").save(os.path.join(save_path, image_name))
def gen_np_array_mem_(self,
results,
classes=cfg.darknet.classes,
det=cfg.xception.det1,
size=cfg.xception.size):
def resize_img(img, size):
img_croped = img.crop(
(-((size - img.size[0]) / 2), -((size - img.size[1]) / 2),
img.size[0] + (size - img.size[0]) / 2,
img.size[1] + (size - img.size[1]) / 2))
img_resized = img_croped.resize((size, size))
return img_resized
tiff_dict = get_tiff_dict()
if self.input_file not in tiff_dict:
raise Exception("XCEPTION PREPROCESS %s NOT FOUND" %
self.input_file)
try:
try:
slide = openslide.OpenSlide(tiff_dict[self.input_file])
except:
slide = TSlide(tiff_dict[self.input_file])
except:
raise Exception('TIFF FILE OPEN FAILED => %s' % self.input_file)
cell_list = []
cell_index = {}
index = 0
for x_y, boxes in results.items():
for box in boxes:
if box[0] in classes and box[1] > det:
# print(x_y)
_, x, y = re.findall(self.pattern, x_y)[0]
# print("1=> %s, %s" % (x, y))
x = int(x) + int(box[2][0])
y = int(y) + int(box[2][1])
w = int(box[2][2])
h = int(box[2][3])
# print("2=> %s, %s" % (x, y))
# center_x = x + int(w / 2 + 0.5)
# center_y = y + int(h / 2 + 0.5)
# w_ = max(w, h)
# h_ = w_
# x_ = center_x - int(w_ / 2 + 0.5)
# y_ = center_y - int(h_ / 2 + 0.5)
# x_ = 0 if x_ < 0 else x_
# y_ = 0 if y_ < 0 else y_
cell = slide.read_region((x, y), 0, (w, h)).convert("RGB")
# cell = slide.read_region((x_, y_), 0, (w_, h_)).convert("RGB")
# image_name = "%s_%s_%s_%s.jpg" % (x, y, w, h)
# cell.save(os.path.join('/home/tsimage/Development/DATA/middle_cells', image_name))
cell_list.append(np.array(resize_img(cell, size)))
cell_index[index] = [x_y, list(box)]
index += 1
slide.close()
# return np.asarray(cell_list), cell_index
return cell_list, cell_index