286 lines
10 KiB
Python
286 lines
10 KiB
Python
"""
|
||
测试随便一份文件结果
|
||
"""
|
||
|
||
import numpy as np
|
||
import torch
|
||
import torch.nn.functional as F
|
||
|
||
from BCGDataset import BCG_Operation
|
||
from Deep_Model import Unet,Fivelayer_Lstm_Unet,Fivelayer_Unet,Sixlayer_Unet
|
||
|
||
modle_dic = {'Fivelayer_Unet':Fivelayer_Unet(),
|
||
'Fivelayer_Lstm_Unet':Fivelayer_Lstm_Unet(),
|
||
'Sixlayer_Unet':Sixlayer_Unet(),
|
||
'U_net':Unet()
|
||
}
|
||
def evaluate(test_data, model,fs,useCPU):
|
||
orgBCG = test_data
|
||
operation = BCG_Operation()
|
||
# 降采样
|
||
orgBCG = operation.down_sample(orgBCG, down_radio=int(fs//100)).copy() #一开始没加.copy()会报错，后来加了就没事了，结果没影响
|
||
# plt.figure()
|
||
# plt.plot(orgBCG)
|
||
# plt.show()
|
||
orgBCG = orgBCG.reshape(-1, 1000)
|
||
# test dataset
|
||
orgData = torch.FloatTensor(orgBCG).unsqueeze(1)
|
||
# predict
|
||
if useCPU == True:
|
||
gpu = False
|
||
else:
|
||
gpu = torch.cuda.is_available()
|
||
if gpu:
|
||
orgData = orgData.cuda()
|
||
model.cuda()
|
||
with torch.no_grad():
|
||
y_hat = model(orgData)
|
||
y_prob = F.sigmoid(y_hat)
|
||
beat = (y_prob>0.5).float().view(-1).cpu().data.numpy()
|
||
beat_diff = np.diff(beat)
|
||
up_index = np.argwhere(beat_diff==1)
|
||
down_index = np.argwhere(beat_diff==-1)
|
||
|
||
return beat,up_index,down_index,y_prob
|
||
|
||
def find_TPeak(data,peaks,th=50):
|
||
"""
|
||
找出真实的J峰或R峰
|
||
:param data: BCG或ECG数据
|
||
:param peaks: 初步峰值（从label中导出的location_R）
|
||
:param th: 范围阈值
|
||
:return: 真实峰值
|
||
"""
|
||
return_peak = []
|
||
for peak in peaks:
|
||
if peak>len(data):continue
|
||
min_win,max_win = max(0,int(peak-th)),min(len(data),int(peak+th))
|
||
return_peak.append(np.argmax(data[min_win:max_win])+min_win)
|
||
return return_peak
|
||
|
||
def new_calculate_beat(y,predict,th=0.5,up=10,th1=100,th2=45): #通过预测计算回原来J峰的坐标 输入：y_prob,predict=ture,up*10,降采样多少就乘多少
|
||
"""
|
||
加上不应期算法，消除误判的峰
|
||
:param y: 预测输出值或者标签值（label）
|
||
:param predict: ture or false
|
||
:param up: 降采样为多少就多少
|
||
:return: 预测的J峰位置
|
||
"""
|
||
if predict:
|
||
beat = np.where(y>th,1,0)
|
||
else:
|
||
beat = y
|
||
beat_diff = np.diff(beat) #一阶差分
|
||
up_index = np.argwhere(beat_diff == 1).reshape(-1)
|
||
down_index = np.argwhere(beat_diff == -1).reshape(-1)
|
||
# print(up_index,down_index)
|
||
# print(y)
|
||
|
||
# print(y[up_index[4]+1:down_index[4]+1])
|
||
|
||
if len(up_index)==0:
|
||
return [0]
|
||
if up_index[0] > down_index[0]:
|
||
down_index = np.delete(down_index, 0)
|
||
if up_index[-1] > down_index[-1]:
|
||
up_index = np.delete(up_index, -1)
|
||
|
||
"""
|
||
加上若大于130点都没有一个心跳时，降低阈值重新判决一次，一般降到0.3就可以了；； 但是对于体动片段降低阈值可能又会造成误判，而且出现体动的话会被丢弃，间隔时间也长
|
||
"""
|
||
# print("初始：",up_index.shape,down_index.shape)
|
||
i = 0
|
||
lenth1 = len(up_index)
|
||
while i < len(up_index)-1:
|
||
if abs(up_index[i+1]-up_index[i]) > th1:
|
||
re_prob = y[down_index[i]+15:up_index[i+1]-15] #原本按正常应该是两个都+1的，但是由于Unet输出低于0.6时，把阈值调小后会在附近一两个点也变为1，会影响判断
|
||
# print(re_prob.shape)
|
||
beat1 = np.where(re_prob > 0.1, 1, 0)
|
||
# print(beat1)
|
||
if sum(beat1) != 0 and beat1[0] != 1 and beat1[-1] != 1:
|
||
insert_up_index,insert_down_index = add_beat(re_prob,th=0.1)
|
||
# print(insert_up_index,insert_down_index,i)
|
||
if len(insert_up_index) > 1:
|
||
l = i+1
|
||
|
||
for u,d in zip(insert_up_index,insert_down_index):
|
||
up_index = np.insert(up_index,l,u+down_index[i]+1+15) #np.insert(arr, obj, values, axis) arr原始数组，可一可多，obj插入元素位置，values是插入内容，axis是按行按列插入。
|
||
down_index = np.insert(down_index,l,d+down_index[i]+1+15)
|
||
l = l+1
|
||
# print('l=', l)
|
||
elif len(insert_up_index) == 1:
|
||
# print(i)
|
||
up_index = np.insert(up_index,i+1,down_index[i]+insert_up_index+1+15)
|
||
down_index = np.insert(down_index,i+1,down_index[i]+insert_down_index+1+15)
|
||
i = i + len(insert_up_index) + 1
|
||
else:
|
||
i = i+1
|
||
continue
|
||
else:
|
||
i = i+1
|
||
# print("最终：",up_index.shape,down_index.shape)
|
||
|
||
"""
|
||
添加不应期
|
||
"""
|
||
new_up_index = up_index
|
||
new_down_index = down_index
|
||
flag = 0
|
||
i = 0
|
||
lenth = len(up_index)
|
||
while i < lenth:
|
||
if abs(up_index[i+1]-up_index[i]) < th2:
|
||
prob_forward = y[up_index[i]+1:down_index[i]+1]
|
||
prob_backward = y[up_index[i+1]+1:down_index[i+1]+1]
|
||
|
||
forward_score = 0
|
||
back_score = 0
|
||
|
||
forward_count = down_index[i] - up_index[i]
|
||
back_count = down_index[i+1] - up_index[i+1]
|
||
|
||
forward_max = np.max(prob_forward)
|
||
back_max = np.max(prob_backward)
|
||
|
||
forward_min = np.min(prob_forward)
|
||
back_min = np.min(prob_backward)
|
||
|
||
forward_average = np.mean(prob_forward)
|
||
back_average = np.mean(prob_backward)
|
||
|
||
if forward_count > back_count:
|
||
forward_score = forward_score + 1
|
||
else:back_score = back_score + 1
|
||
|
||
if forward_max > back_max:
|
||
forward_score = forward_score + 1
|
||
else:back_score = back_score + 1
|
||
|
||
if forward_min < back_min:
|
||
forward_score = forward_score + 1
|
||
else:back_score = back_score + 1
|
||
|
||
if forward_average > back_average:
|
||
forward_score = forward_score + 1
|
||
else:back_score = back_score + 1
|
||
|
||
if forward_score >=3:
|
||
up_index = np.delete(up_index, i+1)
|
||
down_index = np.delete(down_index, i+1)
|
||
flag = 1
|
||
elif back_score >=3:
|
||
up_index = np.delete(up_index, i)
|
||
down_index = np.delete(down_index, i)
|
||
flag = 1
|
||
elif forward_score == back_score:
|
||
if forward_average > back_average:
|
||
up_index = np.delete(up_index, i + 1)
|
||
down_index = np.delete(down_index, i + 1)
|
||
flag = 1
|
||
else:
|
||
up_index = np.delete(up_index, i)
|
||
down_index = np.delete(down_index, i)
|
||
flag = 1
|
||
if flag == 1:
|
||
i = i
|
||
flag = 0
|
||
else: i = i+1
|
||
|
||
else:i = i + 1
|
||
|
||
if i > len(up_index)-2:
|
||
break
|
||
# elif abs(up_index[i+1]-up_index[i]) > 120:
|
||
# print("全部处理之后",up_index.shape,down_index.shape)
|
||
predict_J = (up_index.reshape(-1) + down_index.reshape(-1)) // 2*up
|
||
# predict_J = predict_J.astype(int)
|
||
|
||
return predict_J
|
||
|
||
def add_beat(y,th=0.2): #通过预测计算回原来J峰的坐标 输入：y_prob,predict=ture,up*10,降采样多少就乘多少
|
||
"""
|
||
:param y: 预测输出值或者标签值（label）
|
||
:param predict: ture or false
|
||
:param up: 降采样为多少就多少
|
||
:return: 预测的J峰位置
|
||
"""
|
||
beat1 = np.where(y>th,1,0)
|
||
beat_diff1 = np.diff(beat1) #一阶差分
|
||
add_up_index = np.argwhere(beat_diff1 == 1).reshape(-1)
|
||
add_down_index = np.argwhere(beat_diff1 == -1).reshape(-1)
|
||
# print(beat1)
|
||
# print(add_up_index,add_down_index)
|
||
if len(add_up_index) > 0:
|
||
if add_up_index[0] > add_down_index[0]:
|
||
add_down_index = np.delete(add_down_index, 0)
|
||
if add_up_index[-1] > add_down_index[-1]:
|
||
add_up_index = np.delete(add_up_index, -1)
|
||
return add_up_index, add_down_index
|
||
else:
|
||
return 0
|
||
|
||
def calculate_beat(y,predict,th=0.5,up=10): #通过预测计算回原来J峰的坐标 输入：y_prob,predict=ture,up*10,降采样多少就乘多少
|
||
"""
|
||
:param y: 预测输出值或者标签值（label）
|
||
:param predict: ture or false
|
||
:param up: 降采样为多少就多少
|
||
:return: 预测的J峰位置
|
||
"""
|
||
if predict:
|
||
beat = np.where(y>th,1,0)
|
||
else:
|
||
beat = y
|
||
beat_diff = np.diff(beat) #一阶差分
|
||
up_index = np.argwhere(beat_diff == 1).reshape(-1)
|
||
down_index = np.argwhere(beat_diff == -1).reshape(-1)
|
||
if len(up_index)==0:
|
||
return [0]
|
||
if up_index[0] > down_index[0]:
|
||
down_index = np.delete(down_index, 0)
|
||
if up_index[-1] > down_index[-1]:
|
||
up_index = np.delete(up_index, -1)
|
||
predict_J = (up_index.reshape(-1) + down_index.reshape(-1)) // 2*up
|
||
# predict_J = predict_J.astype(int)
|
||
|
||
return predict_J
|
||
|
||
def Jpeak_Detection(bcg_data, detector_method, fs, low_cut, high_cut, th1, th2, th3, th4, useCPU):
|
||
if detector_method == "Fivelayer_Unet_1":
|
||
molde_name = "Fivelayer_Unet"
|
||
filename = 1
|
||
elif detector_method == "Fivelayer_Unet_2":
|
||
molde_name = "Fivelayer_Unet"
|
||
filename = 2
|
||
elif detector_method == "Fivelayer_Lstm_Unet_1":
|
||
molde_name = "Fivelayer_Lstm_Unet"
|
||
filename = 1
|
||
elif detector_method == "Fivelayer_Lstm_Unet_2":
|
||
molde_name = "Fivelayer_Lstm_Unet"
|
||
filename = 2
|
||
|
||
model_dir = "./result/" + molde_name + '/' + str(filename) + ".pkl" # 模型路径
|
||
# preprocessing = BCG_Operation(sample_rate=1000)
|
||
# BCG = preprocessing.down_sample(BCG, down_radio=int(1000 / fs))
|
||
|
||
bcg_data = bcg_data[:len(bcg_data) // (fs * 10) * fs * 10]
|
||
|
||
model = modle_dic[molde_name]
|
||
model.load_state_dict(torch.load(model_dir, map_location=torch.device('cpu')))
|
||
model.eval()
|
||
preprocessing = BCG_Operation(sample_rate=fs)
|
||
bcg = preprocessing.Butterworth(bcg_data, "bandpass", low_cut=low_cut, high_cut=high_cut, order=3) * th4
|
||
|
||
# J峰预测
|
||
beat, up_index, down_index, y_prob = evaluate(bcg, model=model, fs=fs, useCPU=useCPU)
|
||
y_prob = y_prob.cpu().reshape(-1).data.numpy()
|
||
|
||
predict_J = new_calculate_beat(y_prob, 1, th=0.6, up=fs // 100, th1=th1, th2=th2)
|
||
|
||
predict_J = find_TPeak(bcg, predict_J, th=int(th3 * fs / 1000))
|
||
predict_J = np.array(predict_J)
|
||
|
||
Interval = np.full(len(bcg), np.nan)
|
||
for i in range(len(predict_J) - 1):
|
||
Interval[predict_J[i]: predict_J[i + 1]] = predict_J[i + 1] - predict_J[i]
|
||
|
||
return bcg, predict_J, Interval |