癫痫脑电信号CNN分类实战 二

癫痫脑电信号CNN分类实战 二

本次CNN分类实战项目没有采用Python的生物信号提取包MNE

主要内容
1数据提取
原始数据

def createSpectrogram(data, S=0):
    global nSpectogram
    global signalsBlock
    global inB
    signals=np.zeros((22,59,114))
    
    t=0
    movement=int(S*256)
    if(S==0):
        movement=_SIZE_WINDOW_SPECTOGRAM        
    while data.shape[1]-(t*movement+_SIZE_WINDOW_SPECTOGRAM) > 0:
        # CREAZIONE DELLO SPETROGRAMMA PER TUTTI I CANALI
        for i in range(0, 22):
            start = t*movement
            stop = start+_SIZE_WINDOW_SPECTOGRAM
            signals[i,:]=createSpec(data[i,start:stop])
        if(signalsBlock is None):
            signalsBlock=np.array([signals])
        else:
            signalsBlock=np.append(signalsBlock, [signals], axis=0)
        nSpectogram=nSpectogram+1
        if(signalsBlock.shape[0]==50):
            saveSignalsOnDisk(signalsBlock, nSpectogram)
            signalsBlock=None
            # SALVATAGGIO DI SIGNALS  
        t = t+1
    return (data.shape[1]-t*_SIZE_WINDOW_SPECTOGRAM)*-1

def createSpec(data):
    fs=256
    lowcut=117
    highcut=123

    y=butter_bandstop_filter(data, lowcut, highcut, fs, order=6)
    lowcut=57
    highcut=63
    y=butter_bandstop_filter(y, lowcut, highcut, fs, order=6)
    
    cutoff=1
    y=butter_highpass_filter(y, cutoff, fs, order=6)
    
    Pxx=signal.spectrogram(y, nfft=256, fs=256, return_onesided=True, noverlap=128)[2]    
    Pxx = np.delete(Pxx, np.s_[117:123+1], axis=0)
    Pxx = np.delete(Pxx, np.s_[57:63+1], axis=0)
    Pxx = np.delete(Pxx, 0, axis=0)
    
    result=(10*np.log10(np.transpose(Pxx))-(10*np.log10(np.transpose(Pxx))).min())/(10*np.log10(np.transpose(Pxx))).ptp()
    return result

def createSpectrogram(data, S=0):
    global nSpectogram
    global signalsBlock
    global inB
    signals=np.zeros((22,59,114))
    
    t=0
    movement=int(S*256)
    if(S==0):
        movement=_SIZE_WINDOW_SPECTOGRAM        
    while data.shape[1]-(t*movement+_SIZE_WINDOW_SPECTOGRAM) > 0:
        # CREAZIONE DELLO SPETROGRAMMA PER TUTTI I CANALI
        for i in range(0, 22):
            start = t*movement
            stop = start+_SIZE_WINDOW_SPECTOGRAM
            signals[i,:]=createSpec(data[i,start:stop])
        if(signalsBlock is None):
            signalsBlock=np.array([signals])
        else:
            signalsBlock=np.append(signalsBlock, [signals], axis=0)
        nSpectogram=nSpectogram+1
        if(signalsBlock.shape[0]==50):
            saveSignalsOnDisk(signalsBlock, nSpectogram)
            signalsBlock=None
            # SALVATAGGIO DI SIGNALS  
        t = t+1
    return (data.shape[1]-t*_SIZE_WINDOW_SPECTOGRAM)*-1

def createArrayIntervalData(fSummary):
    preictalInteval=[]
    interictalInterval=[]
    interictalInterval.append(PreIntData(datetime.min, datetime.max))
    files=[]
    firstTime=True
    oldTime=datetime.min # equivalente di 0 nelle date
    startTime=0
    line=fSummary.readline()
    endS=datetime.min
    while(line):
        data=line.split(':')
        if(data[0]=="File Name"):
            nF=data[1].strip()
            s=getTime((fSummary.readline().split(": "))[1].strip())
            if(firstTime):
                interictalInterval[0].start=s
                firstTime=False
                startTime=s
            while s<oldTime:#se cambia di giorno aggiungo 24 ore alla data
                s=s+ timedelta(hours=24)
            oldTime=s
            endTimeFile=getTime((fSummary.readline().split(": "))[1].strip())
            while endTimeFile<oldTime:#se cambia di giorno aggiungo 24 ore alla data
                endTimeFile=endTimeFile+ timedelta(hours=24)
            oldTime=endTimeFile
            files.append(FileData(s, endTimeFile,nF))
            for j in range(0, int((fSummary.readline()).split(':')[1])):
                secSt=int(fSummary.readline().split(': ')[1].split(' ')[0])
                secEn=int(fSummary.readline().split(': ')[1].split(' ')[0])
                ss=s+timedelta(seconds=secSt)- timedelta(minutes=_MINUTES_OF_DATA_BETWEEN_PRE_AND_SEIZURE+_MINUTES_OF_PREICTAL)
                if((len(preictalInteval)==0 or ss > endS) and ss-startTime>timedelta(minutes=20)):
                    ee=ss+ timedelta(minutes=_MINUTES_OF_PREICTAL) 
                    preictalInteval.append(PreIntData(ss,ee))
                endS=s+timedelta(seconds=secEn)
                ss=s+timedelta(seconds=secSt)- timedelta(hours=4) 
                ee=s+timedelta(seconds=secEn)+ timedelta(hours=4) 
                if(interictalInterval[len(interictalInterval)-1].start<ss and interictalInterval[len(interictalInterval)-1].end>ee):
                    interictalInterval[len(interictalInterval)-1].end=ss
                    interictalInterval.append(PreIntData(ee, datetime.max))
                else:
                    if(interictalInterval[len(interictalInterval)-1].start<ee):
                        interictalInterval[len(interictalInterval)-1].start=ee
        line=fSummary.readline()
    fSummary.close()
    interictalInterval[len(interictalInterval)-1].end=endTimeFile
    return preictalInteval, interictalInterval, files

2 滤波过滤

# Filtro taglia banda
def butter_bandstop_filter(data, lowcut, highcut, fs, order):
    nyq = 0.5 * fs
    low = lowcut / nyq
    high = highcut / nyq

    i, u = butter(order, [low, high], btype='bandstop')
    y = lfilter(i, u, data)
    return y

def butter_highpass_filter(data, cutoff, fs, order=5):
    nyq = 0.5 * fs
    normal_cutoff = cutoff / nyq
    b, a = butter(order, normal_cutoff, btype='high', analog=False)
    y = lfilter(b, a, data)
    return y

3划分训练集和测试集

def generate_arrays_for_training(indexPat, paths, start=0, end=100):
    while True:
        from_=int(len(paths)/100*start)
        to_=int(len(paths)/100*end)
        for i in range(from_, int(to_)):
            f=paths[i]
            x = np.load(PathSpectogramFolder+f)
            x=np.array([x])
            x=x.swapaxes(0,1)
            if('P' in f):
                y = np.repeat([[0,1]],x.shape[0], axis=0)
            else:
                y =np.repeat([[1,0]],x.shape[0], axis=0)
            yield(x,y)
            
def generate_arrays_for_predict(indexPat, paths, start=0, end=100):
    while True:
        from_=int(len(paths)/100*start)
        to_=int(len(paths)/100*end)
        for i in range(from_, int(to_)):
            f=paths[i]
            x = np.load(PathSpectogramFolder+f)
            x=np.array([x])
            x=x.swapaxes(0,1)
            yield(x)

2 CNN建模

def createModel():
    input_shape=(1, 22, 59, 114)
    model = Sequential()
    #C1
    model.add(Conv3D(16, (22, 5, 5), strides=(1, 2, 2), padding='valid',activation='relu',data_format= "channels_first", input_shape=input_shape))
    model.add(keras.layers.MaxPooling3D(pool_size=(1, 2, 2),data_format= "channels_first",  padding='same'))
    model.add(BatchNormalization())
    
    #C2
    model.add(Conv3D(32, (1, 3, 3), strides=(1, 1,1), padding='valid',data_format= "channels_first",  activation='relu'))#incertezza se togliere padding
    model.add(keras.layers.MaxPooling3D(pool_size=(1,2, 2),data_format= "channels_first", ))
    model.add(BatchNormalization())
    
    #C3
    model.add(Conv3D(64, (1,3, 3), strides=(1, 1,1), padding='valid',data_format= "channels_first",  activation='relu'))#incertezza se togliere padding
    model.add(keras.layers.MaxPooling3D(pool_size=(1,2, 2),data_format= "channels_first", ))
    model.add(BatchNormalization())
    
    model.add(Flatten())
    model.add(Dropout(0.5))
    model.add(Dense(256, activation='sigmoid'))
    model.add(Dropout(0.5))
    model.add(Dense(2, activation='softmax'))
    
    opt_adam = keras.optimizers.Adam(lr=0.00001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
    model.compile(loss='categorical_crossentropy', optimizer=opt_adam, metrics=['accuracy'])
    
    return model

def main():
    print("START")
    if not os.path.exists(OutputPathModels):
        os.makedirs(OutputPathModels)
    loadParametersFromFile("PARAMETERS_CNN.txt")
    #callback=EarlyStopping(monitor='val_acc', min_delta=0, patience=0, verbose=0, mode='auto', baseline=None)
    callback=EarlyStoppingByLossVal(monitor='val_acc', value=0.975, verbose=1, lower=False)
    print("Parameters loaded")
    
    for indexPat in range(0, len(patients)):
        print('Patient '+patients[indexPat])
        if not os.path.exists(OutputPathModels+"ModelPat"+patients[indexPat]+"/"):
            os.makedirs(OutputPathModels+"ModelPat"+patients[indexPat]+"/")
        loadSpectogramData(indexPat) 
        print('Spectograms data loaded')
        
        result='Patient '+patients[indexPat]+'\n'     
        result='Out Seizure, True Positive, False Positive, False negative, Second of Inter in Test, Sensitivity, FPR \n'
        for i in range(0, nSeizure):
            print('SEIZURE OUT: '+str(i+1))
            
            print('Training start')  
            model = createModel()
            filesPath=getFilesPathWithoutSeizure(i, indexPat)
            
            model.fit_generator(generate_arrays_for_training(indexPat, filesPath, end=75), #end=75),#It take the first 75%
                                validation_data=generate_arrays_for_training(indexPat, filesPath, start=75),#start=75), #It take the last 25%
                                #steps_per_epoch=10000, epochs=10)
                                steps_per_epoch=int((len(filesPath)-int(len(filesPath)/100*25))),#*25), 
                                validation_steps=int((len(filesPath)-int(len(filesPath)/100*75))),#*75),
                                verbose=2,
                                epochs=300, max_queue_size=2, shuffle=True, callbacks=[callback])# 100 epochs è meglio #aggiungere criterio di stop in base accuratezza
            print('Training end')
            
            print('Testing start')
            filesPath=interictalSpectograms[i]
            interPrediction=model.predict_generator(generate_arrays_for_predict(indexPat, filesPath), max_queue_size=4, steps=len(filesPath))
            filesPath=preictalRealSpectograms[i]
            preictPrediction=model.predict_generator(generate_arrays_for_predict(indexPat, filesPath), max_queue_size=4, steps=len(filesPath))
            print('Testing end')
            

            # Creates a HDF5 file 
            model.save(OutputPathModels+"ModelPat"+patients[indexPat]+"/"+'ModelOutSeizure'+str(i+1)+'.h5')
            print("Model saved")
            
            #to plot the model
            #plot_model(model, to_file="CNNModel", show_shapes=True, show_layer_names=True)
            
            if not os.path.exists(OutputPathModels+"OutputTest"+"/"):
                os.makedirs(OutputPathModels+"OutputTest"+"/")
            np.savetxt(OutputPathModels+"OutputTest"+"/"+"Int_"+patients[indexPat]+"_"+str(i+1)+".csv", interPrediction, delimiter=",")
            np.savetxt(OutputPathModels+"OutputTest"+"/"+"Pre_"+patients[indexPat]+"_"+str(i+1)+".csv", preictPrediction, delimiter=",")
            
            secondsInterictalInTest=len(interictalSpectograms[i])*50*30#50 spectograms for file, 30 seconds for each spectogram
            acc=0#accumulator
            fp=0
            tp=0
            fn=0
            lastTenResult=list()
            
            for el in interPrediction:
                if(el[1]>0.5):
                    acc=acc+1
                    lastTenResult.append(1)
                else:
                    lastTenResult.append(0)
                if(len(lastTenResult)>10):
                    acc=acc-lastTenResult.pop(0)
                if(acc>=8):
                  fp=fp+1
                  lastTenResult=list()
                  acc=0
            
            lastTenResult=list()
            for el in preictPrediction:
                if(el[1]>0.5):
                    acc=acc+1
                    lastTenResult.append(1)
                else:
                    lastTenResult.append(0)
                if(len(lastTenResult)>10):
                    acc=acc-lastTenResult.pop(0)
                if(acc>=8):
                  tp=tp+1 
                else:
                    if(len(lastTenResult)==10):
                       fn=fn+1 
                       
            sensitivity=tp/(tp+fn)
            FPR=fp/(secondsInterictalInTest/(60*60))
            
            result=result+str(i+1)+','+str(tp)+','+str(fp)+','+str(fn)+','+str(secondsInterictalInTest)+','
            result=result+str(sensitivity)+','+str(FPR)+'\n'
            print('True Positive, False Positive, False negative, Second of Inter in Test, Sensitivity, FPR')
            print(str(tp)+','+str(fp)+','+str(fn)+','+str(secondsInterictalInTest)+','+str(sensitivity)+','+str(FPR))
        with open(OutputPath, "a+") as myfile:
            myfile.write(result)

项目训练所需的环境
python3.6.5
tensorflow1.10
cuda10.0.130
cudnn7.6.5

显卡2080TI

需要代码直接私信我