SVS格式WSI图像分割

仅仅提供学习使用

---

程序

将原始的coco数据集，标注为矩形框，通过openslide和numpy等工具包进行图片切割。原始的图像大小不规则，100000x100000像素左右，运行下面程序即可得到2000x2000的，尺寸修改、路径等需要自行调整。

import openslide
from openslide.deepzoom import DeepZoomGenerator
import numpy as np
import json
from tqdm import tqdm
import multiprocessing
from concurrent.futures import ThreadPoolExecutor
import threading
import multiprocessing
label = open(r'D:\迅雷下载\核变裂象\mitos_wsi_ccmct_heael.json','r')
label = json.load(label)
IMAGE_SIZE = (2000,2000)
CLASS = label['categories']
# print(CLASS)
# print(label.keys())
# print('总共%d张图像，第一张图片信息'%len(label['images']),label['images'][0])  #32张图片
# print('总共%d个类别，第一个类别信息：'%len(label['categories']),label['categories'][0])  #七个类别
# print('标注数量：',len(label['annotations']))
# print('第一个标注信息',label['annotations'][0])

def get_bbox_class_per_image(label):
    bboxes = [[] for i in range(len(label['images']))]
    image_list = [image['id'] for image in label['images']]

    for label in label['annotations']:
        categories_id = label['category_id']
        if not categories_id:
            continue
        image_id = label['image_id']
        bbox = label['bbox']
        bbox.append(categories_id)
        bboxes[image_list.index(image_id)].append(bbox)
    bboxes = [np.array(bbox,dtype='int') for bbox in bboxes]  # length = 32
    return bboxes
# print(get_bbox_class_per_image(label))

def get_image_label(label,box):

    image_list = [image['id'] for image in label['images']]
    for image, id in zip(label['images'], image_list):
        if id != image['id']:
            print('图片顺序不对')
            continue
        path = r'C:\Users\DELL\Desktop\data' + '\\' + image['file_name']
        slide = openslide.open_slide(path)
        w, h = slide.level_dimensions[0]
        print('strat id:%d image'%image_list.index(id))
        with tqdm(total=h // IMAGE_SIZE[1]*w // IMAGE_SIZE[0], desc=f'process {image_list.index(id)/len(image_list)}', postfix=dict, mininterval=0.3) as pbar:
            for i in range(w // IMAGE_SIZE[0]):
                for j in range(h // IMAGE_SIZE[1]):
                    # box_ = []  #wmin,hmin,wmax,hmax
                    mask_minx = box[image_list.index(id)][:, 2] >= i * 2000+25
                    mask_maxx = box[image_list.index(id)][:, 0] <= (i + 1) * 2000-25
                    mask_miny = box[image_list.index(id)][:, 3] >= j * 2000+25
                    mask_maxy = box[image_list.index(id)][:, 1] <= (j + 1) * 2000-25
                    mask_x_or = np.logical_and(mask_minx,mask_maxx)
                    mask_y_or = np.logical_and(mask_miny,mask_maxy)
                    mask = np.logical_and(mask_x_or,mask_y_or)
                    if not sum(mask):
                        continue

                    box_ = box[image_list.index(id)][mask]
                    box_[:,0:4:2] -= i*2000
                    box_[:,1:4:2] -= j*2000
                    box_ = np.clip(box_,0,2000)
                    np.savetxt('./output/ann/%s_%s_%s.txt'%(str(id-1),str(i),str(j)), box_, fmt='%d')
                    img = slide.read_region((i * 2000, j * 2000), 0, (2000, 2000))
                    img.save('./output/img/%s_%s_%s.png'%(str(id-1),str(i),str(j)))
                    pbar.update(1)
            print('finish id:%d image' % image_list.index(id))
            print('\n')
    return 0

if __name__ == '__main__':
    bboxes = get_bbox_class_per_image(label)
    # print(np.max(bboxes[13],axis=0))  #wmin,hmin,wmax,hmax
    print(len(bboxes[0]))
    get_image_label(label,bboxes)

下面是数据验证的程序以及数据验证的结果：

import cv2
import numpy as np
import matplotlib.pyplot as plt

ann = r'D:\PycharmProjects\simple_skills\切割图像\output\ann\0_9_25.txt'
img = r'D:\PycharmProjects\simple_skills\切割图像\output\img\0_9_25.png'

with open(ann,'r') as f:
    a = f.readlines()
    # for i in a:
    #     print(i.rstrip().split(' '))
    a = [i.rstrip().split(' ') for i in a]
    a = np.array(a,dtype='int')

img = plt.imread(img)*255
img = img.astype('int')
for i in a:
    cv2.rectangle(img,i[:2],i[2:4],(0,255,0),2)
plt.imshow(img)
plt.show()

总结

老老实实的迈好每一步，相信成功就在你面前 甘愿为理想“头破血流”

附录

多线程操作

import openslide
from openslide.deepzoom import DeepZoomGenerator
import numpy as np
import json
from tqdm import tqdm
import multiprocessing
from concurrent.futures import ThreadPoolExecutor
import threading
import multiprocessing
# label = open(r'D:\迅雷下载\核变裂象\mitos_wsi_ccmct_heael.json','r')
# label = json.load(label)
# IMAGE_SIZE = (scale,scale)
# CLASS = label['categories']
# print(CLASS)
# print(label.keys())
# print('总共%d张图像，第一张图片信息'%len(label['images']),label['images'][0])  #32张图片
# print('总共%d个类别，第一个类别信息：'%len(label['categories']),label['categories'][0])  #七个类别
# print('标注数量：',len(label['annotations']))
# print('第一个标注信息',label['annotations'][0])

def get_bbox_class_per_image(label):
    bboxes = [[] for i in range(len(label['images']))]
    image_list = [image['id'] for image in label['images']]

    for label in label['annotations']:
        categories_id = label['category_id']
        if not categories_id:
            continue
        image_id = label['image_id']
        bbox = label['bbox']
        bbox.append(categories_id)
        bboxes[image_list.index(image_id)].append(bbox)
    bboxes = [np.array(bbox,dtype='int') for bbox in bboxes]  # length = 32
    return bboxes
# print(get_bbox_class_per_image(label))

def get_image_label(image,id,image_list,box,IMAGE_SIZE,scale):
    path = r'D:\data' + '\\' + image['file_name']
    slide = openslide.open_slide(path)
    w, h = slide.level_dimensions[0]
    print(w,h)
    print('strat id:%d image'%image_list.index(id))
    with tqdm(total=h // IMAGE_SIZE[1]*w // IMAGE_SIZE[0], desc=f'process {image_list.index(id)/len(image_list)}', postfix=dict, mininterval=0.3) as pbar:
        for i in range(w // IMAGE_SIZE[0]):
            for j in range(h // IMAGE_SIZE[1]):
                # box_ = []  #wmin,hmin,wmax,hmax
                mask_minx = box[image_list.index(id)][:, 2] >= i * scale+25
                mask_maxx = box[image_list.index(id)][:, 0] <= (i + 1) * scale-25
                mask_miny = box[image_list.index(id)][:, 3] >= j * scale+25
                mask_maxy = box[image_list.index(id)][:, 1] <= (j + 1) * scale-25
                mask_x_or = np.logical_and(mask_minx,mask_maxx)
                mask_y_or = np.logical_and(mask_miny,mask_maxy)
                mask = np.logical_and(mask_x_or,mask_y_or)
                if not sum(mask):
                    pbar.update(1)
                    continue

                box_ = box[image_list.index(id)][mask]
                box_[:,0:4:2] -= i*scale
                box_[:,1:4:2] -= j*scale
                box_ = np.clip(box_,0,scale)
                np.savetxt(r'D:\output\Annotations_txt_mel\%s_%s_%s.txt'%(str(id-1),str(i),str(j)), box_, fmt='%d')
                img = slide.read_region((i * scale, j * scale), 0, (scale, scale))
                img.save(r'D:\output\img_mel\%s_%s_%s.png'%(str(id-1),str(i),str(j)))
                pbar.update(1)
        print('finish id:%d image' % image_list.index(id))
        print('\n')

def print_error(value):
    print("error: ", value)

if __name__ == '__main__':
    print('开始运行主线程')
    multiprocessing.freeze_support()
    multiprocessing.Process()
    pool = multiprocessing.Pool(multiprocessing.cpu_count())
    label = open(r'D:\迅雷下载\核变裂象\.mitos_wsi_ccmct_meljson', 'r')
    label = json.load(label)
    IMAGE_SIZE = (640, 640)
    CLASS = label['categories']
    scale = 640
    bboxes = get_bbox_class_per_image(label)
    # print(np.max(bboxes[13],axis=0))  #wmin,hmin,wmax,hmax
    print(len(bboxes[0]))
    # get_image_label(label,bboxes)

    image_list = [image['id'] for image in label['images']]
    print('开始运行子线程')
    for image, id in zip(label['images'], image_list):
        if id != image['id']:
            print('图片顺序不对')
            continue

        pool.apply_async(func=get_image_label,args=[image,id,image_list,bboxes,IMAGE_SIZE,scale,],callback=print_error)
    pool.close()
    pool.join()
    print('主线程运行结束')