小程序开发公司网上海百度移动关键词排名优化

AI医学影像分割方法与流程–基于低场磁共振影像的病灶识别

– 作者:coder_fang
AI框架：PaddleSeg

1. 数据准备，使用MedicalLabelMe进行dcm文件标注，产生同名.json文件。

2. 编写程序生成训练集图片，包括掩码图。

3. 代码如下:

def doCvt(labelDir):outdir=labelDir+"\\output";if os.path.exists(outdir):shutil.rmtree(outdir)os.makedirs(outdir+"\\images") #原始图os.makedirs(outdir + "\\masks1") #可视掩码图,只包含0,255os.makedirs(outdir + "\\masks2") #训练掩码图，只包含0,1start=0files=glob.glob(labelDir+'\\*.json')for f in files:fname=os.path.basename(f)jsondata = json.load(open(labelDir+"\\"+ fname, encoding="utf-8"))imagePath = os.path.join(labelDir, jsondata.get("image_path"))ds = pydicom.read_file(imagePath)  # 读取.dcm文件img = ds.pixel_array  # 提取图像信息img = (img * (255/img.max())).astype(np.uint8)pathout ="{}\\images\\{:05d}.png".format(outdir,start)result=cv2.bilateralFilter(img,11,64,128) #对低场图片进行去噪声cv2.imwrite(pathout, result)label_name_to_value = {'_background_': 0,'H':1}for shape in jsondata.get("frames")[0].get("shapes"):label_name = shape['label']if label_name in label_name_to_value:label_value = label_name_to_value[label_name]else:label_value = len(label_name_to_value)label_name_to_value[label_name] = label_valuemask = np.zeros(img.shape ,dtype=np.uint8)  # 假设图片是灰度图，调整dtype为np.uint8# 遍历所有形状并绘制到掩码上for shape in jsondata.get("frames")[0].get("shapes"):# 获取形状的类型和坐标点shape_type = shape['shape_type']points = shape['points']if shape_type == 'polygon':# 绘制多边形掩码cv2.fillPoly(mask, [np.array(points).reshape((-1, 1, 2)).astype(np.int32)], 1)elif shape_type == 'rectangle':# 绘制矩形掩码（转换为多边形）x, y, w, h = pointspts = np.array([[x, y], [x + w, y], [x + w, y + h], [x, y + h]])cv2.fillPoly(mask, [pts.reshape((-1, 1, 2)).astype(np.int32)], 1)# 可以根据需要添加其他形状的处理方式，如圆形等。# 保存可视掩码图像（例如PNG格式）pathout = "{}\\masks1\\{:05d}.png".format(outdir, start)cv2.imwrite(pathout, (mask * 255).astype(np.uint8))# 保存训练掩码图像（例如PNG格式）pathout = "{}\\masks2\\{:05d}.png".format(outdir, start)cv2.imwrite(pathout, mask.astype(np.uint8))lbl_viz = imgviz.label2rgb(mask, img)start+=1

效果如图:

4. 使用PaddlePaddle 2.6.2版本，单独下载PaddleSeg2.8.1版本进行编译，2.8以前版本会有问题。
5. 生成训练数据和验证数据集对应列表：

def create_list(data_path):image_path = os.path.join(data_path, 'images')label_path = os.path.join(data_path, 'masks2')data_names = os.listdir(image_path)random.shuffle(data_names)  # 打乱数据with open(os.path.join(data_path, 'train_list.txt'), 'w') as tf:with open(os.path.join(data_path, 'val_list.txt'), 'w') as vf:for idx, data_name in enumerate(data_names):img = os.path.join('images', data_name)lab = os.path.join('masks2', data_name)if idx % 9 == 0:  # 90%的作为训练集vf.write(img + ' ' + lab + '\n')else:tf.write(img + ' ' + lab + '\n')print('数据列表生成完成')data_path = 'work/output'
create_list(data_path)  # 生成数据列表

6. 基于列表生成数据集并进行验证:

# 构建训练集
train_transforms = [T.RandomHorizontalFlip(),#水平翻转    T.RandomVerticalFlip(),T.RandomNoise(),T.Resize(),T.Normalize()  # 归一化
]
train_dataset = Dataset(transforms=train_transforms,dataset_root='work/output',num_classes=2,img_channels=1,mode='train',train_path='work/output/train_list.txt',separator=' ',
)
# 构建验证集
val_transforms = [T.Resize(),T.Normalize()
]
val_dataset = Dataset(transforms=val_transforms,dataset_root='work/output',num_classes=2,img_channels=1,mode='val',val_path='work/output/val_list.txt',separator=' ',
)#随机抽取训练集数据进行可视化验证
import matplotlib.pyplot as plt
import numpy as np
plt.figure(figsize=(10,16))
for i in range(1,6,2): _,img,label,_= train_dataset[random.randint(0,100)].items()print(img[1].shape)img = np.transpose(img[1], (1,2,0))img = img*0.5 + 0.5plt.subplot(3,2,i),plt.imshow(img,'gray'),plt.title('img'),plt.xticks([]),plt.yticks([])plt.subplot(3,2,i+1),plt.imshow(label[1],'gray'),plt.title('label'),plt.xticks([]),plt.yticks([])plt.show

7. 构建训练网络并训练:

#可以尝试PaddelSeg中多种分割模型，这里目前测试的BiSeNetV2均衡了效果和训练验证速度，表现还可以
model = BiSeNetV2(num_classes=2,#二分类分割in_channels=1,#单通道lambd=0.25,align_corners=False,pretrained='OutDir/best_model/model.pdparams')base_lr = 0.0005
lr = paddle.optimizer.lr.PolynomialDecay(learning_rate=base_lr, decay_steps=1500, verbose=False)
optimizer = paddle.optimizer.Momentum(lr, parameters=model.parameters(), momentum=0.9, weight_decay=4.0e-
losses = {}
losses['types'] = [CrossEntropyLoss()] *5
losses['coef'] = [0.8,0.2]*5from paddleseg.core import train
train(model=model,train_dataset=train_dataset,val_dataset=val_dataset,optimizer=optimizer,save_dir='OutDir',iters=3000,batch_size=8,save_interval=200,log_iters=20,num_workers=0,losses=losses,use_vdl=True)#开始训练，这里最好使用GPU训练

8. 验证数据，并将预测结果融合到原始图中

#测试结果 transforms = T.Compose([T.Resize(target_size=(512, 512)),T.Normalize()
])#显示图片时调整合适的窗宽窗位
def wwwl(im,window_width,window_level):# 计算最大值和最小值min_val = window_level - window_width / 2max_val = window_level + window_width / 2           #im = ((im - min_val) / (max_val - min_val)) * 255im=np.clip(im,min_val,max_val).astype(np.uint8)return im#加载模型
model = BiSeNetV2(num_classes=2,    in_channels=1,)model_path = 'OutDir/best_model/model.pdparams'
para_state_dict = paddle.load(model_path)
model.set_dict(para_state_dict)import matplotlib.pyplot as plt
import numpy as np#testimg里放置一些病灶图片，尽量不是训练集里数据
files = glob.glob('work/testimg/*.png')
row=(int)(len(files))
plt.figure(figsize=(16,row*8))
i=1;
for f in files:#读取数据    im = cv2.imread(f, 0).astype('float32')im=F.normalize(im,0.5,0.5)h,w = im.shapedata = np.expand_dims(im,axis=0).repeat(1,axis=0)data = data[np.newaxis, ...]data = paddle.to_tensor(data)predata=model(data)output = predata[0].numpy()output = np.argmax(output,axis=1)output = np.squeeze(output)#输入模型预测的时候缩放到(512,512),现在要缩放到原始数据的大小output = cv2.resize(output, (h, w), interpolation=cv2.INTER_NEAREST)output = output.astype(np.uint8)*128#归到255im=(im*0.5+0.5)*255im=wwwl(im,128,60)max_val = np.max(im)im=im*(255/max_val)imgori=cv2.cvtColor(im.astype(np.uint8), cv2.COLOR_GRAY2BGR);imgpre=cv2.cvtColor(output, cv2.COLOR_GRAY2BGR);imgpre[:,:,1]= 0imgpre[:,:,2]= 0mask_img=cv2.addWeighted(imgpre, 0.2, imgori, 0.5, 0)plt.subplot(row,2,i),plt.imshow(mask_img,'gray'),plt.title('img'),plt.xticks([]),plt.yticks([])i+=1plt.show

9. 基于上述模型持续优化并训练，如需模型工程化，可使用paddleseg 中的export.py ,将模型进行转换，转换后可将其集成到python或c++工程使用，转换命令如下：

python export.py --model_path OutDir/best_model/model.pdparams --input_shape 1 1 512 512 --save_dir OutDir/Saved --config work/config.yml 

完！