不常用 - xml工具 - 《深度学习trick》

# @Time    : 2019.10.18
# @Author  : Bohrium.Kwong
# @Licence : bio-totem
import os
import shutil
import lxml.etree as ET
from PIL import ImageDraw, Image
import numpy as np
import copy
import cv2
import glob
import openslide_utils
import sys
sys.path.append('../')
from openslide_utils import Slide
# from xml_generate import get_contours
import matplotlib.pyplot as plt
current_path = os.path.dirname(__file__)
def dist(a, b):
    return round(abs(a[0]-b[0]) + abs(a[1]-b[1]))
def floatrange(start,end,step):
    floatrange_list = []
    for i in range(int((end-start)/step)):
        floatrange_list.append(round(start+step*i,1))
    return floatrange_list
# xml_ori_path = 'where xml files save'
def xml_change_color(xml_ori_path):
    for xml in glob.glob(xml_ori_path+'/*.xml*'):
        #print(xml)
        tree = ET.parse(xml)
        for a in tree.findall('.//Annotation'):
            # print(a.attrib['LineColor'])
            a.attrib['LineColor'] = '65535'
            tree.write(xml,pretty_print=True)
##解析xml文件获得坐标点
def xml_to_region(xml_file):
    """
    parse XML label file and get the points
    :param xml_file: xml file
    :return: region list,region_class
    """
    tree = ET.parse(xml_file)
    region_list = []
    region_class = []
    for color in tree.findall('.//Annotation'):
        if color.attrib['LineColor'] in ['65280']:  #zheli shi biaoqian zhong de biaozhu xian tiao de yanse,'65280'
            # '65280'是绿色,'255'是红色,可以根据自己的实际情况更改这个判断条件(或者直接if True)
            for region in color.findall('./Regions/Region'):
                vertex_list = []
                #region.attrib.get('Type')=='0':
                region_class.append(region.attrib.get('Type')) #region_class:leibie dui ying de suo yin
                for vertex in region.findall('.//Vertices/Vertex'):
                    # parse the 'X' and 'Y' for the vertex
                    vertex_list.append(vertex.attrib)
                region_list.append(vertex_list)
    return region_list,region_class
##解析xml文件获得坐标点
def xml_to_region_pred(xml_file):
    """
    parse XML label file and get the points
    :param xml_file: xml file
    :return: region list,region_class
    """
    tree = ET.parse(xml_file)
    region_list = []
    region_class = []
    for color in tree.findall('.//Annotation'):
        if color.attrib['LineColor'] in ['255']:  #zheli shi biaoqian zhong de biaozhu xian tiao de yanse,'65280'
            # '65280'是绿色,'255'是红色,可以根据自己的实际情况更改这个判断条件(或者直接if True)
            for region in color.findall('./Regions/Region'):
                vertex_list = []
                #region.attrib.get('Type')=='0':
                region_class.append(region.attrib.get('Type')) #region_class:leibie dui ying de suo yin
                for vertex in region.findall('.//Vertices/Vertex'):
                    # parse the 'X' and 'Y' for the vertex
                    vertex_list.append(vertex.attrib)
                region_list.append(vertex_list)
    return region_list,region_class
##将标签点（离散）加工成连续点，并将其绘制到原图上
def region_handler(im, region_list,region_class, level_downsample):
    """
    handle region label point to discrete point, and draw the region point to line
    :param im: the image painted in region line
    :param region_list: region list, region point,
                    eg : [[{'X': '27381.168113', 'Y': '37358.653791'}], [{'X': '27381.168113', 'Y': '37358.653791'}]]
    :param region_class : list,keep the value of region.attrib.get('Type') in elements of region list
                    eg : [0,0,0,1,2,3]
    :param level_downsample: slide level down sample
    :return: image painted in region line of numpy array format
    """
    dr = ImageDraw.Draw(im)
    for r_class, region in enumerate(region_list):
        point_list = []
        if region_class[r_class] == '0' or region_class[r_class] == '3':
            for __, point in enumerate(region):
                X, Y = int(float(point['X'])/level_downsample), int(float(point['Y'])/level_downsample)
                point_list.append((X, Y))
#        points_length = len(point_list)
            x_max = max(point_list, key=lambda point: point[0])[0]
            x_min = min(point_list, key=lambda point: point[0])[0]
            y_max = max(point_list, key=lambda point: point[1])[1]
            y_min = min(point_list, key=lambda point: point[1])[1]
            # mislabeled, here checked by x and y coordinate max and min difference
            if (x_max - x_min < 50) or (y_max - y_min < 50): continue
            if region_class[r_class] == '3':
                dr.arc(point_list, 0, 360, fill='#00FF00', width=10) ##huizhi de yanse
            else:
                dr.line(point_list, fill="#00FF00", width=10)  ##huizhi de yanse
    return im
##zhe shi wei le zai yuantu shang tongshi shengcheng biaoqian he yuce xiantiao shiyong de(fei bi xu)
def region_handler_pred(im, region_list, region_class, level_downsample):
    """
    handle region label point to discrete point, and draw the region point to line
    :param im: the image painted in region line
    :param region_list: region list, region point,
                    eg : [[{'X': '27381.168113', 'Y': '37358.653791'}], [{'X': '27381.168113', 'Y': '37358.653791'}]]
    :param region_class : list,keep the value of region.attrib.get('Type') in elements of region list
                    eg : [0,0,0,1,2,3]
    :param level_downsample: slide level down sample
    :return: image painted in region line of numpy array format
    """
    dr = ImageDraw.Draw(im)
    for r_class, region in enumerate(region_list):
        point_list = []
        if region_class[r_class] == '0' or region_class[r_class] == '3':
            for __, point in enumerate(region):
                X, Y = int(float(point['X']) / level_downsample), int(float(point['Y']) / level_downsample)
                point_list.append((X, Y))
            #        points_length = len(point_list)
            x_max = max(point_list, key=lambda point: point[0])[0]
            x_min = min(point_list, key=lambda point: point[0])[0]
            y_max = max(point_list, key=lambda point: point[1])[1]
            y_min = min(point_list, key=lambda point: point[1])[1]
            # mislabeled, here checked by x and y coordinate max and min difference
            if (x_max - x_min < 50) or (y_max - y_min < 50): continue
            if region_class[r_class] == '3':
                dr.arc(point_list, 0, 360, fill='#0000FF', width=0)  ##huizhi de yanse
            else:
                dr.line(point_list, fill="#0000FF", width=10)  ##huizhi de yanse
    return im
def region_binary_image(tile, region_list, region_class, level_downsample):
    """
    convert the region labeled or not by doctor to binary image
    :param tile: a return image based on the method of Slide class object in 'utils.openslide_utils'
    :param region_list: region list, region point,
                    eg : [[{'X': '27381.168113', 'Y': '37358.653791'}], [{'X': '27381.168113', 'Y': '37358.653791'}]]
    :param region_class : list,keep the value of region.attrib.get('Type') in elements of region list
                    eg : [0,0,0,1,2,3]
    :param level_downsample: slide level down sample
    :return: image painted in region line of numpy array format
    """
    im = Image.new(mode="1", size=tile.size)
    dr = ImageDraw.Draw(im)
    regions_list = []
    for r_class, region in enumerate(region_list):
        point_list = []
        if region_class[r_class] == '0':
            for __, point in enumerate(region):
                X, Y = int(float(point['X']) / level_downsample), int(float(point['Y']) / level_downsample)
                point_list.append((X, Y))
            regions_list.append(point_list)
    # 由于医生的标注会出现不连续(非闭合)的情况，导致提取出来的标注坐标列表，会分成多段，比如：
    #            0 (1979, 798) (2144, 1479)
    #            1 (2139, 1483) (2319, 2162)
    #            2 (2308, 2160) (3003, 1646)
    # 正常情况下，标注坐标列表应该收尾闭合(前后坐标一致)，上下列表之间差异应该较大，如：
    #            12 (1177, 2986) (1177, 2986)
    #            13 (1507, 2942) (1507, 2940)
    # 针对上述第一种情况，需要对提出出来的标注坐标列表进行循环判断，对收尾不一而且和其他标注列表的首坐标相对较近的话，进行合并处理
    pin_jie_flag = []  # 存储已经被拼接过的标注坐标列表序号
    single_list = []  # 存储新标注坐标列表的列表
    for j, p_list in enumerate(regions_list):
        if dist(p_list[0], p_list[-1]) < 50 and j not in pin_jie_flag:
            # 如果收尾坐标距离相差在150范围内(曼哈顿距离)，且未成被拼接过，直接认为这个组坐标无须拼接，存储起来
            single_list.append(p_list)
        elif dist(p_list[0], p_list[-1]) > 50 and j not in pin_jie_flag:
            # 如果收尾坐标距离相差在150范围外(曼哈顿距离)，且未成被拼接过，说明这组坐标是残缺非闭合的，需要对其余标注坐标进行新一轮的循环判断
            for j_2, p_list_2 in enumerate(regions_list):
                if j_2 != j and j_2 not in pin_jie_flag:
                    if dist(p_list[-1], p_list_2[0]) < 50:
                        p_list = p_list + p_list_2.copy()
                        pin_jie_flag.append(j_2)
                    elif dist(p_list[0], p_list_2[-1]) < 50:
                        p_list = p_list_2.copy() + p_list
                        pin_jie_flag.append(j_2)
                    elif dist(p_list[-1], p_list_2[-1]) < 50:
                        p_list_2_new = copy.deepcopy(p_list_2)
                        p_list_2_new.reverse()
                        p_list = p_list + p_list_2_new
                        pin_jie_flag.append(j_2)
                    elif dist(p_list[0], p_list_2[0]) < 50:
                        p_list_2_new = copy.deepcopy(p_list_2)
                        p_list_2_new.reverse()
                        p_list = p_list_2_new + p_list
                        pin_jie_flag.append(j_2)
                    # 当这组非闭合的尾坐标和其他组坐标的首坐标接近到一定范围时(距离是150内),就让当前的非闭合的坐标列表和该组坐标列表相加
                    # 处理完毕之后，将该组坐标的序号增加到已拼接坐标的列表中，确保后续循环不会再判断这个列表
            single_list.append(p_list)
    for points in single_list:
        dr.polygon(points, fill="#ffffff")
    # 由于医生的标注除了出现不连续(非闭合)的情况外，还存在多余勾画的情况，对这种情况暂时没有完整的思路予以接近，先用
    # opencv中的开闭操作组合来进行修补
    # kernel size depends on slide size,original is (20,20)
    kernel = np.ones((3, 3), np.uint8)
    filter_matrix = np.array(im).astype(np.uint8)
    filter_matrix = cv2.morphologyEx(filter_matrix, cv2.MORPH_OPEN, kernel)
    filter_matrix = cv2.morphologyEx(filter_matrix, cv2.MORPH_CLOSE, kernel)
    #    plt.imshow(filter_matrix)
    return filter_matrix
def change_xml_line_color(xml_ori_path):
    for xml in glob.glob(xml_ori_path+'/*.xml*'):
        #print(xml)
        tree = ET.parse(xml)
        for a in tree.findall('.//Annotation'):
            # print(a.attrib['LineColor'])
            a.attrib['LineColor'] = '65535'
            tree.write(xml,pretty_print=True)
def save_label_and_pred_pic(svs_path,model_mode,xml_path_label,xml_path_pred,img_name,threshold=0):
    slide = Slide(svs_path)
    tile = slide.get_thumb()  # 获取2级采样下的全片截图
    tile_copy = tile.copy()
    assert os.path.split(xml_path_label)[-1] == os.path.split(xml_path_pred)[-1]
    region_list_label,region_class_label = xml_to_region(xml_path_label)
    region_list_pred, region_class_pred = xml_to_region_pred(xml_path_pred)
    # #在这里使用xml_utils的方法进行指定区域提取(最终返回的是个True False矩阵)
    #region_process_mask = region_binary_image(tile, region_list_label, region_class_label,slide.get_level_downsample())
    # # 根据上述返回的标注坐标列表生成WSI原图2级采样大小的True False矩阵
    ##zhuyi: zheli shi like huizhi shangqu ,suoyi yao yong yici zhu shi yici
    region_xml_label_draw = region_handler(tile, region_list_label, region_class_label,slide.get_level_downsample())
    # sheng cheng label huizhi de tu
    plt.figure(figsize=(12.4,9.6))
    plt.imshow(region_xml_label_draw)
    plt.axis('off')
    plt.margins(0, 0)
    # plt.savefig(r'/media/totem_disk/totem/guozunhu/Project/(wen tai)projectC_cc/backup/(ds:12\18\24)/MainNet10/min_loss_model/paip/train/validation_in_train/label_pic/{}_{}_label_green.png'.format(img_name,model_mode))
    plt.savefig('/media/totem_disk/totem/guozunhu/Project/new_hooknet/test_ndpi_out_he_hn_cc_ce_train/ori_label/{}'.format(img_name))
    plt.close()
    # region_xml_pred_draw = region_handler_pred(tile_copy, region_list_pred, region_class_pred,slide.get_level_downsample())
    # #sheng cheng pred huizhi de tu
    # plt.imshow(region_xml_pred_draw)
    # plt.axis('off')
    # plt.margins(0, 0)
    # # plt.savefig(r'/media/totem_disk/totem/guozunhu/Project/(wen tai)projectC_cc/backup/ds:4\8\12/MainNet10/min_loss/epoch:600/pred_pic/label_and_pred/{}_{}.png'.format(img_name, model_mode))
    # # if not os.path.exists(r':600/validation_set/first_12_next_4/pred_xml_and_mask/arclength<800/{}/prec_pic'.format(threshold)):
    #   # os.makedirs(r'/media/totem_disk/totem/guozunhu/Project/(wen tai)projectC_cc/backup/ds:4\8\12/MainNet10/halfway_hardsample/min_loss/epoch:600/validation_set/first_12_next_4/pred_xml_and_mask/arclength<800/{}/prec_pic'.format(threshold), exist_ok=True)
    # plt.savefig('/media/totem_disk/totem/guozunhu/Project/lp/test_ndpi_out_he_cc/{}_pred'.format(img_name))
    # plt.close()
    # region_xml_pred_draw = region_handler_pred(tile, region_list_pred, region_class_pred,slide.get_level_downsample())
    # #sheng cheng pred huizhi de tu
    # plt.imshow(region_xml_pred_draw)
    #
    # plt.axis('off')
    # plt.margins(0, 0)
    # plt.savefig('/media/totem_disk/totem/guozunhu/Project/lp/cache_ndpi_he_cc/hard_sample/ori_xml+hard_sample/pic/{}.png'.format(img_name))
    # plt.close()
class Region:
    """"
    handle the template xml format file to insert label svs region
    """
    def __init__(self, xml_file):
        parser = ET.XMLParser(remove_blank_text=True)
        if not os.path.isfile(xml_file):
            template = os.path.join(current_path, "template.xml")
            shutil.copy(template, xml_file)
        self._xml_file = xml_file
        self._tree = ET.parse(xml_file, parser)
    def get_region(self, region_id):
        """
        get region by region id
        :param region_id: region id, 0: green, 1: yellow, 2: red, see the template.xml
        :return: the region
        """
        return self._tree.findall(".//Annotation/Regions")[region_id]
    def add(self, region_id, points):
        """
        add one region to the specified region by region id, the added region is ellipse
        and the parameter points is a rectangle bounded by an ellipse
        :param points: list with two element(upper-left, bottom-right), is the rectangle bounded by an ellipse
        :return:
        """
        region = self.get_region(region_id)
        region_num = len(region.findall(".//Region"))
        region_attr = {
            "Id": str(region_num+1),
            "Type": "2",
            "Zoom": "1",
            "Selected": "1",
            "ImageLocation": "",
            "ImageFocus": "0",
            "Length": "80",
            "Area": "400",
            "LengthMicrons": "20",
            "AreaMicrons": "30",
            "Text": "",
            "NegativeROA": "0",
            "InputRegionId": "0",
            "Analyze": "0",
            "DisplayId": "1"
        }
        region_tag = ET.Element("Region", region_attr)
        region.append(region_tag)
        attributes = ET.SubElement(region_tag, "Attributes")
        vertices = ET.Element("Vertices")
        region_tag.append(vertices)
        for point in points:
            # insert point
            ET.SubElement(vertices, "Vertex", attrib=point)
    def save(self):
        """
        save the xml file
        :return:
        """
        self._tree.write(self._xml_file, pretty_print=True)
if __name__ == '__main__':
    if False:
        model_mode = 'min_loss'
        assert model_mode in ['min_loss','best_f1','last']
        svs_path_root = '/media/totem_disk/totem/guozunhu/Project/new_hooknet/data_cc/valid'
        svs_path_list = glob.glob(svs_path_root+'/*.svs')
        # svs_path_list = sorted(glob.glob(svs_path_root+'/*.svs'))
        xml_path_label_root  = r'/media/totem_disk/totem/guozunhu/Project/new_hooknet/data_cc/valid'
        xml_path_pred_root = r'/media/totem_disk/totem/guozunhu/Project/new_hooknet/data_cc/valid'
        for svs_path in svs_path_list:
            # xml_path_label = '/media/totem_disk/totem/guozunhu/Project/lp/cache_ndpi_he_cc/hard_sample/ori_xml+hard_sample'
            # xml_path_pred = '/media/totem_disk/totem/guozunhu/Project/lp/cache_ndpi_he_cc/hard_sample/ori_xml+hard_sample'
            img_name = os.path.split(svs_path)[-1].split('.svs')[0]
            # if 'neg' in img_name:
            #     continue
            xml_path_label = glob.glob(xml_path_label_root+'/*{}.xml'.format(img_name))[0]
            xml_path_pred = glob.glob(xml_path_pred_root+'/*{}.xml'.format(img_name))[0]
            assert os.path.split(xml_path_label)[-1] == os.path.split(xml_path_pred)[-1]
            save_label_and_pred_pic(svs_path,model_mode,xml_path_label,xml_path_pred,img_name)
            print("Finished:{}".format(os.path.split(svs_path)[-1]))
    if True:
        import yaml
        model_mode = 'min_loss'
        for i in range(0,10):
            th = i / 10
            mean_dice = {'mean_dice':0}
            low_dice_list = []
            all_img_path = glob.glob('/media/totem_disk/totem/guozunhu/Project/new_hooknet/test_ndpi_out_he_hn_cc_ce_train/ds:12 20_new_dataset/merge_mask/{}/*.png'.format(th))
            for img_path in all_img_path:
                img_name = os.path.splitext(img_path)[0].split('/')[-1]
                # svs_path = '/media/totem_disk/totem/guozunhu/Project/lp/all_svs/first_batch/all/{}.svs'.format(img_name)
                # xml_path_label = '/media/totem_disk/totem/guozunhu/Project/lp/all_svs/first_batch/all/{}.xml'.format(img_name)
                # region_list, region_class = xml_to_region(xml_path_label)
                # slide = Slide(svs_path)
                # tile = slide.get_thumb()
                # level_downsample = slide.get_level_downsample()
                # label = region_binary_image(tile,region_list,region_class,level_downsample)
                # label = np.where(label < 0.5, 0, label)
                # label = np.where(label >= 0.5, 1, label)
                # cv2.imwrite('/media/totem_disk/totem/guozunhu/Project/lp_2/cache_ndpi_he_cc/hot_pic/true_mask/{}_label.png'.format(img_name),label*255)
                # print(img_name)
                if 'neg' in img_name:
                    continue
                label = cv2.imread('/media/totem_disk/totem/guozunhu/Project/lp/backup/before/all/true_mask/{}_label.png'.format(img_name),0)
                label = np.where(label < 127.5, 0, label)
                label = np.where(label >= 127.5, 1, label)
                # todo:分别计算不同ds倍率下热图的dice
                all_ds = [2,4,6]
                for ds in all_ds:
                    out = cv2.imread('/media/totem_disk/totem/guozunhu/Project/temp_3/cache_ndpi_he_cc/M10:4_8_12(最大像素值)/hot_pic/{}_{}.png'.format(img_name,ds),0)
                    out_2 = out.copy()
                    out_2 = np.where(out_2 < 128,0,out_2)
                    out_2 = np.where(out_2 >= 128,1,out_2)
                    tp = (out_2 * label).sum(dtype=np.int)
                    fp = out_2.sum(dtype=np.int) - tp
                    label_p_num = label.sum(dtype=np.int)
                    dice = 2 * tp / (label_p_num + tp + fp + 1e-8)
                    out = cv2.cvtColor(out, cv2.COLOR_GRAY2BGR)
                    cv2.putText(out, 'dice:{}'.format(dice), (200, 600), cv2.FONT_HERSHEY_SIMPLEX, 20,(255, 255,255), 25, cv2.LINE_AA)
                    cv2.imwrite('/media/totem_disk/totem/guozunhu/Project/temp_3/cache_ndpi_he_cc/M10:4_8_12(最大像素值)/hot_pic/dice/{}_{}.png'.format(img_name,ds),out)
                # todo:计算merge_mask de dice
                out = cv2.imread('/media/totem_disk/totem/guozunhu/Project/lp_3_u2net/predict/merge_mask/{}/{}.png'.format(th,img_name),0)
                out_2 = out.copy()
                out_2 = np.where(out_2 < 128,0,out_2)
                out_2 = np.where(out_2 >= 128,1,out_2)
                tp = (out_2 * label).sum(dtype=np.int)
                fp = out_2.sum(dtype=np.int) - tp
                label_p_num = label.sum(dtype=np.int)
                dice = 2 * tp / (label_p_num + tp + fp + 1e-8)
        #         if dice < 0.7:
        #             low_dice_list.append(img_name)
        #         # mean_dice['mean_dice'] += (dice / len(all_img_path)).item()
        #         mean_dice['mean_dice'] += (dice / 28).item()
        #
        #         out = cv2.cvtColor(out, cv2.COLOR_GRAY2BGR)
        #         cv2.putText(out, 'dice:{}'.format(dice), (200, 200), cv2.FONT_HERSHEY_SIMPLEX, 8,(255, 255,255), 10, cv2.LINE_AA)
        #         save_root_path = '/media/totem_disk/totem/guozunhu/Project/lp_3_u2net/predict/dice' + '/' + str(th)
        #         if not os.path.exists(save_root_path):
        #             os.mkdir(save_root_path)
        #         save_path = save_root_path + '/' + img_name + '.png'
        #         cv2.imwrite(save_path,out)
        #     mean_dice['low_dice_list'] = low_dice_list
        #     yaml.safe_dump(mean_dice,open('/media/totem_disk/totem/guozunhu/Project/lp_3_u2net/predict/dice/{}/mean_dice.yml'.format(th),'w'))
    if False:
        import yaml
        mean_dice = {'mean_dice':0}
        low_dice_list = []
        all_img_path = glob.glob('/media/totem_disk/totem/guozunhu/Project/new_hooknet/test_ndpi_out_he_hn_cc_ce_train/ds:12 20_valid/*_hm_m.png')
        for img_path in all_img_path:
            img_name = os.path.splitext(img_path)[0].split('/')[-1].split('_hm_m')[0]
            # if 'neg' in img_name:
            #     continue
            svs_path = '/media/totem_disk/totem/guozunhu/Project/new_hooknet/data_cc/valid/{}.svs'.format(img_name)
            xml_path_label = '/media/totem_disk/totem/guozunhu/Project/new_hooknet/data_cc/valid/{}.xml'.format(img_name)
            region_list, region_class = xml_to_region(xml_path_label)
            slide = Slide(svs_path)
            tile = slide.get_thumb()
            level_downsample = slide.get_level_downsample()
            label = region_binary_image(tile,region_list,region_class,level_downsample)
            label = np.where(label < 0.5, 0, label)
            label = np.where(label >= 0.5, 1, label)
            print(img_name)
            pred = cv2.imread(img_path,0)
            pred = cv2.resize(pred,(label.shape[1],label.shape[0]))
            pred = np.where(pred < 128,0,pred)
            pred = np.where(pred >= 128,1,pred)
            tp = (pred * label).sum(dtype=np.int)
            fp = pred.sum(dtype=np.int) - tp
            label_p_num = label.sum(dtype=np.int)
            dice = 2 * tp / (label_p_num + tp + fp + 1e-8)
            if dice < 0.9:
                low_dice_list.append(img_name)
            mean_dice['mean_dice'] += (dice / len(all_img_path)).item()
        mean_dice['low_dice_list'] = low_dice_list
        yaml.safe_dump(mean_dice,open('/media/totem_disk/totem/guozunhu/Project/new_hooknet/test_ndpi_out_he_hn_cc_ce_train/ds:12 20_valid/dice/mean_dice.yml','w'))
    # xml_path_label = '/media/totem_disk/totem/guozunhu/Project/cc/xml_check/cc/he/pred_xml/634671.xml'
    # region_list, region_class = xml_to_region(xml_path_label)
    # slide = Slide('/media/totem_disk/totem/guozunhu/Project/lp_3_u2net/backup/valid/634671.svs')
    # tile = slide.get_thumb()
    # level_downsample = slide.get_level_downsample()
    # label = region_binary_image(tile,region_list,region_class,level_downsample)
    # label = np.where(label < 0.5, 0, label)
    # label = np.where(label >= 0.5, 1, label)
    # cv2.imwrite('/media/totem_disk/totem/guozunhu/Project/cc/xml_check/cc/he/pred_xml/634671.png',label*255)
    # print(img_name)
    # if True:
    #     cache_paths = glob.glob('/media/totem_disk/totem/guozunhu/Project/lp_4_hooknet/cache_ndpi_he_cc/*_cache')
    #     for cache_path in cache_paths:
    #         svs_name = os.path.splitext(cache_path)[0].split('/')[-1].split('_cache')[0]
    #         svs_path = '/media/totem_disk/totem/Data_20201014/LABIAL_SALIVARY_GLAND_BIOPSY/{}.svs'.format(svs_name)
    #         slide = Slide(svs_path)
    #         tile = slide.get_thumb()
    #         save_path = '/media/totem_disk/totem/guozunhu/Project/lp_4_hooknet/xml_out/ori/{}.png'.format(svs_name)
    #         if os.path.exists(save_path):
    #             continue
    #         tile.save(save_path)
    #         print(svs_name)