MONAI实现病理影像肿瘤检测全流程

最新推荐文章于 2025-07-23 09:38:25 发布

DeepHacking

最新推荐文章于 2025-07-23 09:38:25 发布

阅读量224

点赞数 4

CC 4.0 BY-SA版权

文章标签： python 深度学习人工智能

本文链接：https://blog.youkuaiyun.com/deephacking/article/details/149494945

MONAI对病理影像大图（WSI）肿瘤检测流程，覆盖：

WSI Patch 采样与数据集构建（用 WSIPatchDataset）
模型训练（以 UNet 为例）
滑动窗口推理（sliding_window_inference）全图分割
可视化和全图掩膜生成

重点使用 MONAI 的功能，兼容 OpenSlide 超大病理图像。代码默认已准备好数据（WSI 图片和掩膜），适合你快速二次开发。

1. 环境准备

pip install monai[all] openslide-python torch torchvision matplotlib

2. Patch 数据集构建（WSIPatchDataset）

假设数据结构如下：

dataset/
  images/
    case1.svs
    case2.svs
    ...
  masks/
    case1_mask.tif
    case2_mask.tif
    ...

Patch 数据集实现

from monai.data import WSIPatchDataset
from monai.transforms import (
    LoadImaged, AddChanneld, ScaleIntensityd, ToTensord, Compose
)
import glob

# 获取 WSI 和掩膜路径
image_paths = sorted(glob.glob('dataset/images/*.svs'))
mask_paths  = sorted(glob.glob('dataset/masks/*.tif'))

# 构建数据对
data_dicts = [{"image": img, "mask": msk} for img, msk in zip(image_paths, mask_paths)]

# Patch 配置
patch_size = (256, 256)
level = 0      # 金字塔层级，通常0为最高分辨率
patch_per_wsi = 1000  # 每张WSI采样patch数量

# 数据增强
train_transforms = Compose([
    LoadImaged(keys=["image", "mask"], reader='WSIReader'),
    AddChanneld(keys=["image", "mask"]),
    ScaleIntensityd(keys=["image"]),
    ToTensord(keys=["image", "mask"])
])

# 构建 WSIPatchDataset
train_ds = WSIPatchDataset(
    data=data_dicts,
    patch_size=patch_size,
    transform=train_transforms,
    level=level,
    patch_per_wsi=patch_per_wsi,
    return_patch_iter=False,   # 直接返回patch张量和标签
)
from torch.utils.data import DataLoader
train_loader = DataLoader(train_ds, batch_size=8, shuffle=True)

3. 定义分割模型

from monai.networks.nets import UNet
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
net = UNet(
    spatial_dims=2,    # 2D patch
    in_channels=1,     # 通道数可根据你的数据调整
    out_channels=2,    # 例如肿瘤/背景二分类
    channels=(32, 64, 128, 256, 512),
    strides=(2, 2, 2, 2),
    num_res_units=2
).to(device)

4. 训练模型

from monai.losses import DiceLoss
import torch.optim as optim

loss_function = DiceLoss(to_onehot_y=True, softmax=True)
optimizer = optim.Adam(net.parameters(), lr=1e-4)

max_epochs = 20

for epoch in range(max_epochs):
    net.train()
    epoch_loss = 0
    for batch_data in train_loader:
        inputs = batch_data["image"].to(device)
        labels = batch_data["mask"].to(device)
        optimizer.zero_grad()
        outputs = net(inputs)
        loss = loss_function(outputs, labels)
        loss.backward()
        optimizer.step()
        epoch_loss += loss.item()
    print(f"Epoch {epoch+1}/{max_epochs}, Loss: {epoch_loss/len(train_loader):.4f}")

5. 滑动窗口推理（全图掩膜重建）

推理时按窗口滑动整张WSI，拼接mask：

from monai.inferers import sliding_window_inference
from monai.data import WSIReader
import numpy as np

# 假设要对 case1.svs 推理
wsi_path = 'dataset/images/case1.svs'

# 1. 读取整张 WSI
reader = WSIReader()
wsi_img = reader.read(wsi_path, level=level)
wsi_arr = np.array(wsi_img)  # shape: (H, W, C)
if wsi_arr.ndim == 3 and wsi_arr.shape[2] > 1:
    wsi_arr = wsi_arr[..., 0]  # 若多通道，仅用一通道或做通道合成
wsi_tensor = torch.tensor(wsi_arr).unsqueeze(0).unsqueeze(0).float().to(device)  # shape: [1,1,H,W]

# 2. 滑动窗口推理
roi_size = (256, 256)
sw_batch_size = 4

net.eval()
with torch.no_grad():
    pred_mask = sliding_window_inference(
        wsi_tensor, roi_size, sw_batch_size, net, overlap=0.25
    )
    pred_mask = torch.argmax(pred_mask, dim=1).cpu().numpy()[0]  # [H,W]

6. 可视化分割结果

import matplotlib.pyplot as plt

plt.figure(figsize=(12,6))
plt.subplot(1,2,1)
plt.title('Original WSI')
plt.imshow(wsi_arr, cmap='gray')
plt.subplot(1,2,2)
plt.title('Predicted Tumor Mask')
plt.imshow(pred_mask, cmap='Reds', alpha=0.7)
plt.show()