本文基于Unet分割网络，利用49例头部磁共振T1、T2数据，通过配准使二者解剖位置一致，转换数据格式并裁剪窗宽窗位，构建数据集。以T1为输入、T2为标签训练Unet进行回归，用SSIM评估，经200轮训练，最佳SSIM达0.571，实现由T1生成虚拟T2图像。
☞☞☞AI 智能聊天, 问答助手, AI 智能搜索, 免费无限量使用 DeepSeek R1 模型☜☜☜
这个想法是基于
[1]高留刚, 李春迎, 陆正大,等. 基于卷积神经网络生成虚拟平扫CT图像[J]. 中国医学影像技术, 2022, 38(3):5.
【1】文章使用Unet对增强CT数据进行训练，最终预测生成对应的虚拟平扫CT图像，达到只需要对患者扫描一次CT即可，避免患者接受过多的放射辐射。
【2】因为本项目没有对应的CT数据，刚好有一个T1、T2磁共振头部公开数据，先磁共振数据进行配准，然后输入T1模态的数据到Unet中然后生成对应的T2模态数据。
数量：49例
模态和部位：磁共振的T1和T2头部数据 
#解压数据!unzip -o /home/aistudio/data/data146796/mydata.zip -d /home/aistudio/work
#antspyx 是一个配准工具包，SImpleITK处理医学数据!pip install antspyx SimpleITK
import osimport antsimport SimpleITK as sitkfrom tqdm import tqdmimport numpy as npimport randomimport matplotlib.pyplot as pltimport paddleimport cv2
#因为配准花费的时间太久太久了，本项目的数据已经包含好了配准后的数据。#因此不用运行此代码path = '/home/aistudio/work/mydata/IXI-T1'for f in tqdm(os.listdir(path)):
f_path = os.path.join(path,f)
m_path = os.path.join('/home/aistudio/work/mydata/IXI-T2',f.replace('T1','T2'))
f_img = ants.image_read(f_path)
m_img = ants.image_read(m_path)
mytx = ants.registration(fixed=f_img, moving=m_img, type_of_transform='SyN') # 将形变场作用于moving图像，得到配准后的图像
warped_img = ants.apply_transforms(fixed=f_img, moving=m_img, transformlist=mytx['fwdtransforms'],
interpolator="linear") # 将配准后图像的direction/origin/spacing和原图保持一致
warped_img.set_direction(f_img.direction)
warped_img.set_origin(f_img.origin)
warped_img.set_spacing(f_img.spacing)
img_name = '/home/aistudio/work/mydata/IXI-T2-warped/'+ f.replace('T1','Warped') # ants.image_write(warped_img, img_name)print("End")原始数据格式是NIFIT，是属于医疗数据格式的一种。把NIFIT数据转换成numpy,并对窗宽窗位进行裁剪到0~255。并生成txt文档，用作构建DataSet使用。
random.seed(2022)
save_path = '/home/aistudio/work/npdata'save_origin_path = "/home/aistudio/work/npdata/origin/"save_label_path = "/home/aistudio/work/npdata/label/"if not os.path.exists(save_path):
os.mkdir(save_path)
os.mkdir(save_origin_path)
os.mkdir(save_label_path)
origin_data = '/home/aistudio/work/mydata/IXI-T1'f_list = os.listdir(origin_data)
random.shuffle(f_list)
split = int(len(f_list)*0.9)
t_f_list = f_list[:split]
v_f_list = f_list[split:]def gen_txt(f_list,name):
txt = open( os.path.join(save_path,name),'w') for f in f_list: if '.ipynb_checkpoints' in f: continue
f_path = os.path.join(origin_data,f)
tag_path = os.path.join(origin_data.replace('T1','T2-warped'),f.replace('T1','Warped'))
f_sitkData = sitk.ReadImage(f_path)
tag_sitkData = sitk.ReadImage(tag_path)
f_npData = sitk.GetArrayFromImage(f_sitkData)
tag_npData = sitk.GetArrayFromImage(tag_sitkData)
f.split('-')[0] for i in range(100,200):
f_slice = np.rot90(f_npData[:,i,:],-1)
tag_slice = np.rot90(tag_npData[:,i,:],-1)
f_slice = (f_slice - 0) / ((835 - 0) / 255)
np.clip(f_slice, 0, 255, out=f_slice)
tag_slice = (tag_slice - 0) / ((626 - 0) / 255)
np.clip(tag_slice, 0, 255, out=tag_slice)
f_save_path = save_origin_path+f.split('-')[0]+'_'+str(i) +'.npy'
tag_save_path = save_label_path+f.split('-')[0]+'_'+str(i) +'.npy'
np.save(f_save_path,f_slice)
np.save(tag_save_path,tag_slice)
txt.write(f_save_path + ' ' + tag_save_path+ '\n')
txt.close()
gen_txt(t_f_list,name='train.txt')
gen_txt(v_f_list,name='val.txt')print('完成')数据增强只采用 缩放到256x256。
from paddle.io import Dataset, DataLoaderfrom paddle.vision.transforms import Resizeclass MyDataset(Dataset):
def __init__(self, data_dir, txt_path, transform=None):
super(MyDataset, self).__init__()
self.data_list = [] with open(txt_path,encoding='utf-8') as f: for line in f.readlines():
image_path, label_path = line.split(' ')
image_path = image_path.strip()
label_path = label_path.strip()
image_path = os.path.join(data_dir, image_path)
label_path = os.path.join(data_dir,label_path)
self.data_list.append([image_path, label_path])
self.transform = transform def __getitem__(self, index):
image_path, label_path = self.data_list[index]
image = np.load(image_path)
label = np.load(label_path) if self.transform is not None:
image = self.transform(image)
label = self.transform(label)
image = image[np.newaxis,:].astype('float32')
label = label[np.newaxis,:].astype('float32') return image, label def __len__(self):
return len(self.data_list)
BatchSize = 24transform = Resize(size=(256,256))
t_dataset = MyDataset('/home/aistudio','work/npdata/train.txt',transform=transform)
v_dataset = MyDataset('/home/aistudio','work/npdata/val.txt',transform=transform)
train_loader = DataLoader(t_dataset,batch_size=BatchSize)
val_loader = DataLoader(v_dataset,batch_size=BatchSize)SSIM (Structure Similarity Index Measure) 结构衡量指标
结构相似指标可以衡量图片的失真程度，也可以衡量两张图片的相似程度，SSIM指标感知模型，即更符合人眼的直观感受。SSIM 主要考量图片的三个关键特征：亮度（Luminance）, 对比度（Contrast）, 结构 (Structure)。SSIM取值约接近1，两张图片的相似度约接近。两张图片一模一样，SSIM=1
参考文章 https://zhuanlan.zhihu.com/p/399215180
def ssim(img1, img2):
C1 = (0.01 * 255)**2
C2 = (0.03 * 255)**2
img1 = img1.astype(np.float64)
img2 = img2.astype(np.float64)
kernel = cv2.getGaussianKernel(11, 1.5)
window = np.outer(kernel, kernel.transpose())
mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] # valid
mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
mu1_sq = mu1**2
mu2_sq = mu2**2
mu1_mu2 = mu1 * mu2
sigma1_sq = cv2.filter2D(img1**2, -1, window)[5:-5, 5:-5] - mu1_sq
sigma2_sq = cv2.filter2D(img2**2, -1, window)[5:-5, 5:-5] - mu2_sq
sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) *
(sigma1_sq + sigma2_sq + C2)) return ssim_map.mean()def calculate_ssim(img1, img2):
if not img1.shape == img2.shape: raise ValueError('Input images must have the same dimensions.') if img1.ndim == 2: return ssim(img1, img2) elif img1.ndim == 3: if img1.shape[2] == 3:
ssims = [] for i in range(3):
ssims.append(ssim(img1, img2)) return np.array(ssims).mean() elif img1.shape[2] == 1: return ssim(np.squeeze(img1), np.squeeze(img2)) else: raise ValueError('Wrong input image dimensions.')for data in t_dataset:
img1,img2 = data
plt.figure(figsize=(10,6))
plt.subplot(1,2,1)
plt.imshow(np.squeeze(img1),'gray')
plt.title("T1 mode")
plt.subplot(1,2,2)
plt.imshow(np.squeeze(img2),'gray')
plt.title("T2 mode") print(f'SSIM:{calculate_ssim(np.squeeze(img1),np.squeeze(img2))}') breakSSIM:0.4927141870957079
<Figure size 720x432 with 2 Axes>
Unet网络是一种语义分割网络，输入原始图像和对应的标签Mask图像到Unet网络中进行分割，最终输出和标签Mask一致的图像。实际也是对像素值进行分类。把原始图像中的像素值进行分类，到底是属于Mask图像中像素值0，还是像素值1等等。这里是使用Unet网络进行回归，因此网络输出的类别是1。
from model.unet import UNet img = paddle.rand([2, 1, 256, 256]) model = UNet() out = model(paddle.rand([2, 1, 256, 256]))print(out.shape)
[2, 1, 256, 256]
def evaluation(model,val_dataset):
"""验证"""
model.eval()
ssims = [] for data in val_dataset:
image,label = data
image = image[np.newaxis,:].astype('float32')
image = paddle.to_tensor(image)
pre = model(image).numpy()
pre = np.squeeze(pre).astype('int32')
ssims.append(calculate_ssim(np.squeeze(label.astype('int32')),pre)) return np.array(ssims).mean()def train(model,epochs):
if not os.path.exists('/home/aistudio/save_model'):
os.mkdir('/home/aistudio/save_model')
steps = int(len(t_dataset)/BatchSize) * epochs
best_ssim = 0 #记录最优的ssim得分
model.train()
scheduler = paddle.optimizer.lr.PolynomialDecay(learning_rate=0.05, decay_steps=steps, verbose=False)
opt = paddle.optimizer.Adam(learning_rate=scheduler, parameters=model.parameters())
for epoch_id in range(epochs): for batch_id, data in enumerate(train_loader()):
images, labels = data
images = paddle.to_tensor(images)
labels = paddle.to_tensor(labels)
predicts = model(images)
mse_loss = paddle.nn.MSELoss()
l1_loss = paddle.nn.L1Loss()
loss = mse_loss(predicts, labels)+l1_loss(predicts,labels) if batch_id % 200 == 0: print("epoch: {}, batch: {}, loss is: {}".format(epoch_id, batch_id, loss.numpy()))
loss.backward()
opt.step()
opt.clear_grad() #训练过程中，保存模型参数
if epoch_id % 10 == 0:
paddle.save(model.state_dict(), '/home/aistudio/save_model/'+str(epoch_id)+'model.pdparams')
model.eval() #训练过程中，预览效果，输入T1图片，预测t2图片
data = t_dataset[150]
img,label = data input = img[np.newaxis,:].astype('float32') input = paddle.to_tensor(input)
pre = model(input).numpy()
pre = np.squeeze(pre)
plt.figure(figsize=(12,6))
plt.subplot(1,3,1)
plt.imshow(np.squeeze(img),'gray')
plt.title("input: T1 mode")
plt.subplot(1,3,2)
plt.imshow(np.squeeze(label),'gray')
plt.title("target:T2 mode")
plt.subplot(1,3,3)
plt.imshow(pre,'gray')
plt.title("pre: T2 mode")
plt.show()
v_ssim_mean = evaluation(model, v_dataset) if v_ssim_mean > best_ssim:
paddle.save(model.state_dict(), '/home/aistudio/save_model/best_model.pdparams')
best_ssim = v_ssim_mean
model.train() print(f'val_SSIM: {v_ssim_mean},Best_SSIM:{best_ssim}')
#启动训练过程train(model,epochs= 200)epoch: 0, batch: 0, loss is: [7316.7188]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.39978080604275146,Best_SSIM:0.39978080604275146 epoch: 1, batch: 0, loss is: [2628.9192] epoch: 2, batch: 0, loss is: [2724.6853] epoch: 3, batch: 0, loss is: [2688.8508] epoch: 4, batch: 0, loss is: [2686.88] epoch: 5, batch: 0, loss is: [2608.186] epoch: 6, batch: 0, loss is: [2520.5952] epoch: 7, batch: 0, loss is: [2692.3694] epoch: 8, batch: 0, loss is: [2702.1028] epoch: 9, batch: 0, loss is: [2628.6016] epoch: 10, batch: 0, loss is: [2485.4382]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4802982366278921,Best_SSIM:0.4802982366278921 epoch: 11, batch: 0, loss is: [2507.683] epoch: 12, batch: 0, loss is: [2434.3452] epoch: 13, batch: 0, loss is: [2408.262] epoch: 14, batch: 0, loss is: [2332.7097] epoch: 15, batch: 0, loss is: [2270.0247] epoch: 16, batch: 0, loss is: [2179.795] epoch: 17, batch: 0, loss is: [2138.3933] epoch: 18, batch: 0, loss is: [2093.4106] epoch: 19, batch: 0, loss is: [1887.1506] epoch: 20, batch: 0, loss is: [2024.5569]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.49954443953370115,Best_SSIM:0.49954443953370115 epoch: 21, batch: 0, loss is: [1973.8619] epoch: 22, batch: 0, loss is: [1956.6097] epoch: 23, batch: 0, loss is: [1909.3981] epoch: 24, batch: 0, loss is: [1875.7285] epoch: 25, batch: 0, loss is: [1837.553] epoch: 26, batch: 0, loss is: [1812.5168] epoch: 27, batch: 0, loss is: [1785.5492] epoch: 28, batch: 0, loss is: [1774.7295] epoch: 29, batch: 0, loss is: [1752.758] epoch: 30, batch: 0, loss is: [1756.8203]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5426780015544808,Best_SSIM:0.5426780015544808 epoch: 31, batch: 0, loss is: [1735.7894] epoch: 32, batch: 0, loss is: [1827.0814] epoch: 33, batch: 0, loss is: [1753.9219] epoch: 34, batch: 0, loss is: [1779.7365] epoch: 35, batch: 0, loss is: [1752.4994] epoch: 36, batch: 0, loss is: [1747.4668] epoch: 37, batch: 0, loss is: [1735.7844] epoch: 38, batch: 0, loss is: [1769.8733] epoch: 39, batch: 0, loss is: [1759.1631] epoch: 40, batch: 0, loss is: [1630.8364]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4373114577520071,Best_SSIM:0.5426780015544808 epoch: 41, batch: 0, loss is: [1611.3628] epoch: 42, batch: 0, loss is: [1715.6615] epoch: 43, batch: 0, loss is: [1763.4269] epoch: 44, batch: 0, loss is: [1737.9359] epoch: 45, batch: 0, loss is: [1772.5739] epoch: 46, batch: 0, loss is: [1738.1847] epoch: 47, batch: 0, loss is: [1775.006] epoch: 48, batch: 0, loss is: [1911.8531] epoch: 49, batch: 0, loss is: [2033.4453] epoch: 50, batch: 0, loss is: [1987.3456]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5579299673856584,Best_SSIM:0.5579299673856584 epoch: 51, batch: 0, loss is: [1836.5175] epoch: 52, batch: 0, loss is: [1861.1411] epoch: 53, batch: 0, loss is: [1845.1718] epoch: 54, batch: 0, loss is: [1740.5963] epoch: 55, batch: 0, loss is: [1660.3125] epoch: 56, batch: 0, loss is: [1726.9702] epoch: 57, batch: 0, loss is: [1635.8535] epoch: 58, batch: 0, loss is: [1610.6799] epoch: 59, batch: 0, loss is: [1568.5394] epoch: 60, batch: 0, loss is: [1617.876]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5634165131362056,Best_SSIM:0.5634165131362056 epoch: 61, batch: 0, loss is: [1561.1827] epoch: 62, batch: 0, loss is: [1597.7261] epoch: 63, batch: 0, loss is: [1525.599] epoch: 64, batch: 0, loss is: [1548.1874] epoch: 65, batch: 0, loss is: [1488.4364] epoch: 66, batch: 0, loss is: [1513.0366] epoch: 67, batch: 0, loss is: [1505.5546] epoch: 68, batch: 0, loss is: [1516.5098] epoch: 69, batch: 0, loss is: [1467.4851] epoch: 70, batch: 0, loss is: [1585.8317]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5709971328311436,Best_SSIM:0.5709971328311436 epoch: 71, batch: 0, loss is: [1457.6887] epoch: 72, batch: 0, loss is: [1467.6174] epoch: 73, batch: 0, loss is: [1471.8882] epoch: 74, batch: 0, loss is: [1459.0536] epoch: 75, batch: 0, loss is: [1458.9404] epoch: 76, batch: 0, loss is: [1634.24] epoch: 77, batch: 0, loss is: [1650.2292] epoch: 78, batch: 0, loss is: [1437.5302] epoch: 79, batch: 0, loss is: [1421.7593] epoch: 80, batch: 0, loss is: [1413.1609]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5701560896501561,Best_SSIM:0.5709971328311436 epoch: 81, batch: 0, loss is: [1409.3071] epoch: 82, batch: 0, loss is: [1404.0553] epoch: 83, batch: 0, loss is: [1420.5449] epoch: 84, batch: 0, loss is: [1398.2864] epoch: 85, batch: 0, loss is: [1488.5781] epoch: 86, batch: 0, loss is: [1401.356] epoch: 87, batch: 0, loss is: [1410.7175] epoch: 88, batch: 0, loss is: [1500.6272] epoch: 89, batch: 0, loss is: [1416.3497] epoch: 90, batch: 0, loss is: [1380.4263]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5465621052241267,Best_SSIM:0.5709971328311436 epoch: 91, batch: 0, loss is: [1372.7605] epoch: 92, batch: 0, loss is: [1367.5848] epoch: 93, batch: 0, loss is: [1376.7378] epoch: 94, batch: 0, loss is: [1370.2089] epoch: 95, batch: 0, loss is: [1365.8433] epoch: 96, batch: 0, loss is: [1369.1512] epoch: 97, batch: 0, loss is: [1369.3212] epoch: 98, batch: 0, loss is: [1404.3115] epoch: 99, batch: 0, loss is: [1354.4286] epoch: 100, batch: 0, loss is: [1382.9532]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5378666148453135,Best_SSIM:0.5709971328311436 epoch: 101, batch: 0, loss is: [1387.7471] epoch: 102, batch: 0, loss is: [1424.3396] epoch: 103, batch: 0, loss is: [1349.6024] epoch: 104, batch: 0, loss is: [1366.5181] epoch: 105, batch: 0, loss is: [1352.3512] epoch: 106, batch: 0, loss is: [1353.1865] epoch: 107, batch: 0, loss is: [1352.0631] epoch: 108, batch: 0, loss is: [1343.2163] epoch: 109, batch: 0, loss is: [1345.3413] epoch: 110, batch: 0, loss is: [1346.6847]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5586893791021966,Best_SSIM:0.5709971328311436 epoch: 111, batch: 0, loss is: [1343.9454] epoch: 112, batch: 0, loss is: [1349.0774] epoch: 113, batch: 0, loss is: [1337.5864] epoch: 114, batch: 0, loss is: [1429.8538] epoch: 115, batch: 0, loss is: [1349.8104] epoch: 116, batch: 0, loss is: [1335.1368] epoch: 117, batch: 0, loss is: [1337.9221] epoch: 118, batch: 0, loss is: [1352.1951] epoch: 119, batch: 0, loss is: [1347.4542] epoch: 120, batch: 0, loss is: [1336.236]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4977443843057142,Best_SSIM:0.5709971328311436 epoch: 121, batch: 0, loss is: [1341.1677] epoch: 122, batch: 0, loss is: [1325.7648] epoch: 123, batch: 0, loss is: [1322.5] epoch: 124, batch: 0, loss is: [1322.8683] epoch: 125, batch: 0, loss is: [1319.6675] epoch: 126, batch: 0, loss is: [1319.3303] epoch: 127, batch: 0, loss is: [1320.0753] epoch: 128, batch: 0, loss is: [1316.222] epoch: 129, batch: 0, loss is: [1336.7915] epoch: 130, batch: 0, loss is: [1321.2736]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.47030362428033146,Best_SSIM:0.5709971328311436 epoch: 131, batch: 0, loss is: [1339.2891] epoch: 132, batch: 0, loss is: [1318.2183] epoch: 133, batch: 0, loss is: [1325.5737] epoch: 134, batch: 0, loss is: [1308.886] epoch: 135, batch: 0, loss is: [1298.8752] epoch: 136, batch: 0, loss is: [1302.0266] epoch: 137, batch: 0, loss is: [1279.0824] epoch: 138, batch: 0, loss is: [1313.5026] epoch: 139, batch: 0, loss is: [1317.8335] epoch: 140, batch: 0, loss is: [1325.1211]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.5129117285623543,Best_SSIM:0.5709971328311436 epoch: 141, batch: 0, loss is: [1307.1959] epoch: 142, batch: 0, loss is: [1312.2686] epoch: 143, batch: 0, loss is: [1320.1404] epoch: 144, batch: 0, loss is: [1326.3783] epoch: 145, batch: 0, loss is: [1308.4757] epoch: 146, batch: 0, loss is: [1318.8088] epoch: 147, batch: 0, loss is: [1297.8518] epoch: 148, batch: 0, loss is: [1296.301] epoch: 149, batch: 0, loss is: [1301.2318] epoch: 150, batch: 0, loss is: [1290.0143]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4650640222375466,Best_SSIM:0.5709971328311436 epoch: 151, batch: 0, loss is: [1320.3531] epoch: 152, batch: 0, loss is: [1311.9688] epoch: 153, batch: 0, loss is: [1302.6729] epoch: 154, batch: 0, loss is: [1302.4635] epoch: 155, batch: 0, loss is: [1286.9762] epoch: 156, batch: 0, loss is: [1284.86] epoch: 157, batch: 0, loss is: [1290.937] epoch: 158, batch: 0, loss is: [1274.3252] epoch: 159, batch: 0, loss is: [1280.6571] epoch: 160, batch: 0, loss is: [1286.4685]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4561319579342975,Best_SSIM:0.5709971328311436 epoch: 161, batch: 0, loss is: [1270.9604] epoch: 162, batch: 0, loss is: [1287.5771] epoch: 163, batch: 0, loss is: [1269.8563] epoch: 164, batch: 0, loss is: [1286.9226] epoch: 165, batch: 0, loss is: [1272.0933] epoch: 166, batch: 0, loss is: [1274.8392] epoch: 167, batch: 0, loss is: [1272.4178] epoch: 168, batch: 0, loss is: [1266.7671] epoch: 169, batch: 0, loss is: [1282.4298] epoch: 170, batch: 0, loss is: [1247.015]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.49646587443554935,Best_SSIM:0.5709971328311436 epoch: 171, batch: 0, loss is: [1250.8077] epoch: 172, batch: 0, loss is: [1287.3835] epoch: 173, batch: 0, loss is: [1248.3782] epoch: 174, batch: 0, loss is: [1249.7322] epoch: 175, batch: 0, loss is: [1273.4828] epoch: 176, batch: 0, loss is: [1253.0632] epoch: 177, batch: 0, loss is: [1281.6831] epoch: 178, batch: 0, loss is: [1280.6669] epoch: 179, batch: 0, loss is: [1245.0092] epoch: 180, batch: 0, loss is: [1257.4398]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.46512517332255776,Best_SSIM:0.5709971328311436 epoch: 181, batch: 0, loss is: [1242.4584] epoch: 182, batch: 0, loss is: [1254.164] epoch: 183, batch: 0, loss is: [1238.1926] epoch: 184, batch: 0, loss is: [1243.1198] epoch: 185, batch: 0, loss is: [1248.0677] epoch: 186, batch: 0, loss is: [1240.6064] epoch: 187, batch: 0, loss is: [1252.3843] epoch: 188, batch: 0, loss is: [1230.179] epoch: 189, batch: 0, loss is: [1237.1627] epoch: 190, batch: 0, loss is: [1232.43]
<Figure size 864x432 with 3 Axes>
val_SSIM: 0.4939679476601408,Best_SSIM:0.5709971328311436 epoch: 191, batch: 0, loss is: [1245.5471] epoch: 192, batch: 0, loss is: [1255.1567] epoch: 193, batch: 0, loss is: [1250.8704] epoch: 194, batch: 0, loss is: [1223.4425] epoch: 195, batch: 0, loss is: [1224.8658] epoch: 196, batch: 0, loss is: [1230.7142] epoch: 197, batch: 0, loss is: [1229.7573] epoch: 198, batch: 0, loss is: [1235.9465] epoch: 199, batch: 0, loss is: [1225.0635]
以上就是基于分割网络Unet生成虚拟图像的详细内容，更多请关注php中文网其它相关文章！
每个人都需要一台速度更快、更稳定的 PC。随着时间的推移，垃圾文件、旧注册表数据和不必要的后台进程会占用资源并降低性能。幸运的是，许多工具可以让 Windows 保持平稳运行。
广告
收藏
收藏
收藏
Copyright 2014-2025 https://www.php.cn/ All Rights Reserved | php.cn | 湘ICP备2023035733号
833
