第二課：肺部X光結果的語義分割

本課主要介紹應用MD.ai使用U-Net來進行語義分割；

使用MD.ai註釋器檢視DICOM影象，並建立影象級別註釋。然後使用MD.ai python客戶端庫下載影象和註釋，準備資料集，然後用於訓練模型進行分類。

MD.ai官網如下：https://www.md.ai/

是一個專門用於醫療AI的開源庫，不過目前好像就只有這個kaggle專案呀，可以很方便的看這次比賽裡面的各個圖片。。

參照第一課安裝和匯入需要的庫並進行初始準備：

這裡需要匯入一個token，是mdai需要的，進入https://public.md.ai/hub/settings#tokens

p = mdai_client.project('aGq4k6NW', path='') 

該行程式碼用於在網上搜索為aGq4k6NW的專案，path表示下載路徑，置空表示當前路徑。

pip install pydicom 
pip install tqdm 
pip install imgaug
pip install mdai

import os
import sys
import random
import math
import numpy as np
import cv2
import matplotlib.pyplot as plt
import json
import pydicom
from imgaug import augmenters as iaa

import skimage.io
import skimage.measure
from tqdm import tqdm
from PIL import Image

import requests
import shutil
import zipfile

import mdai

mdai_client = mdai.Client(domain='public.md.ai', access_token="")

p = mdai_client.project('aGq4k6NW', path='./lesson2-data')

p.show_label_groups()

labels_dict = {'L_A8Jm3d':1 # Lung   
              }

print(labels_dict)
p.set_labels_dict(labels_dict)

p.show_datasets() 

dataset = p.get_dataset_by_id('D_rQLwzo')
dataset.prepare()

image_ids = dataset.get_image_ids()
len(image_ids)

# 檢視部分訓練圖片 
mdai.visualize.display_images(image_ids[:3], cols=2)
 

下載UNet實現：

簡要介紹一下UNet:

UNET_URL = 'https://s3.amazonaws.com/md.ai-ml-lessons/unet.zip'
UNET_ZIPPED = 'unet.zip'

if not os.path.exists(UNET_ZIPPED): 
    r = requests.get(UNET_URL, stream=True)
    if r.status_code == requests.codes.ok:
        with open(UNET_ZIPPED, "wb") as f:
            shutil.copyfileobj(r.raw, f)
    else:
        r.raise_for_status()

    with zipfile.ZipFile(UNET_ZIPPED) as zf:
        zf.extractall()
 

進行訓練

imgs_anns_dict = dataset.imgs_anns_dict

from unet import dataset
from unet import dilated_unet
from unet import train

images, masks = dataset.load_images(imgs_anns_dict)

img_index = random.choice(range(len(imgs_anns_dict)))

print(img_index)
img_fps = list(imgs_anns_dict.keys())
img_fp = img_fps[img_index]
img = mdai.visualize.load_dicom_image(img_fp)
ann = imgs_anns_dict[img_fp]
img_width = img.shape[1]
img_height = img.shape[0]

mask = np.zeros((img_height, img_width), dtype=np.uint8) 
for a in ann:     
    vertices = np.array(a['data']['vertices'])
    vertices = vertices.reshape((-1,2))                     
    cv2.fillPoly(mask, np.int32([vertices]), (255,255,255))
    
plt.figure(figsize=(30, 20))
plt.subplot(2,3,1)
plt.imshow(img, cmap=plt.cm.bone)
plt.axis('off')

plt.subplot(2,3,2)
plt.imshow(mask, cmap=plt.cm.bone)
plt.axis('off')

plt.subplot(2,3,3)              
plt.imshow(cv2.bitwise_and(img, img, mask=mask.astype(np.uint8)), cmap=plt.cm.bone)
plt.axis('off')

plt.subplot(2,3,4)
plt.imshow(images[img_index,:,:,0], cmap=plt.cm.bone)
plt.axis('off')

plt.subplot(2,3,5)
plt.imshow(masks[img_index,:,:,0], cmap=plt.cm.bone)
plt.axis('off')

plt.subplot(2,3,6)
plt.imshow(cv2.bitwise_and(images[img_index,:,:,0], images[img_index,:,:,0], 
                           mask=masks[img_index,:,:,0].astype(np.uint8)), cmap=plt.cm.bone)
plt.axis('off')

import tensorflow as tf 
config = tf.ConfigProto()
config.gpu_options.allow_growth = True

CONFIG_FP = 'unet/configs/11.json'
name = os.path.basename(CONFIG_FP).split('.')[0]
print(name)

with open(CONFIG_FP, 'r') as f:
    config = json.load(f)

# increase the number of epochs for better prediction 
history = train.train(config, name, images,masks, num_epochs=20)

#畫出學習曲線

import matplotlib.pyplot as plt

print(history.history.keys())

plt.figure()
plt.plot(history.history['acc'], 'orange', label='Training accuracy')
plt.plot(history.history['val_acc'], 'blue', label='Validation accuracy')
plt.plot(history.history['loss'], 'red', label='Training loss')
plt.plot(history.history['val_loss'], 'green', label='Validation loss')
plt.legend()
plt.show()

from keras.models import load_model
import keras.backend as K

model_name = 'unet/trained/model_'+name+'.hdf5'
print(model_name)
model = load_model(model_name, custom_objects={'dice': train.dice, 'iou': train.iou})

images, masks = dataset.load_images(imgs_anns_dict)

plt.figure(figsize=(20, 10))

img_index = random.choice(range(len(images)))

plt.subplot(1,4,1)
random_img = images[img_index,:,:,0]
plt.imshow(random_img, cmap=plt.cm.bone)
plt.axis('off')
plt.title('Lung X-Ray')

plt.subplot(1,4,2)
random_mask = masks[img_index,:,:,0]
plt.imshow(random_mask, cmap=plt.cm.bone)
plt.axis('off')
plt.title('Mask Ground Truth')

random_img_2 = np.expand_dims(np.expand_dims(random_img, axis=0), axis=3)
mask = model.predict(random_img_2)[0][:,:,0] > 0.5
plt.subplot(1,4,3)
plt.imshow(mask, cmap=plt.cm.bone)
plt.axis('off')
plt.title('Predicted Mask')

plt.subplot(1,4,4)
plt.imshow(cv2.bitwise_and(random_img, random_img, mask=mask.astype(np.uint8)), cmap=plt.cm.bone)
plt.axis('off')
plt.title('Predicted Lung Segmentation')

預測出的肺部區域