def logisticRegressionArtificialData(dataFolder):
uniformReflectances = np.load(dataFolder + "2015February1704:00PMreflectancesRandom1000000photons.npy")
uniformReflectances = mch.normalizeImageQuotient(uniformReflectances, iqBand=1)
gaussReflectances = np.load(dataFolder + "2015February1703:56PMreflectancesRandomNonUniform1000000photons.npy")
gaussReflectances = mch.normalizeImageQuotient(gaussReflectances, iqBand=1)
return uniformReflectances, gaussReflectances
def perfect(dataFolder):
trainingParameters = np.load(dataFolder + "2015February1612:04AMparamtersPerfectArtificial.npy")
trainingParameters = trainingParameters[:,0:2] # only BVF and Vs for perfect data simulations
trainingReflectances = np.load(dataFolder + "2015February1612:04AMreflectancesPerfectArtificial1000000photons.npy")
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances, iqBand=1)
testParameters = np.load(dataFolder + "2015February1702:02AMparamtersRandom.npy")
testParameters = testParameters[:,0:2] # only BVF and Vs for perfect data simulations
testReflectances = np.load(dataFolder + "2015February1702:02AMreflectancesRandom.npy")
testReflectances = mch.normalizeImageQuotient(testReflectances, iqBand=1)
return trainingParameters, trainingReflectances, testParameters, testReflectances
def noisy(dataFolder):
trainingParameters = np.load(dataFolder + "2015February2208:16PMNoisyRandomTraining10000parameters.npy")
trainingParameters = trainingParameters[:, 0:2]
trainingReflectances = np.load(
dataFolder + "2015February2208:16PMNoisyRandomTrainingreflectances1000000photons.npy"
)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
testParameters = np.load(dataFolder + "2015February2301:23AMNoisyRandomTesting10000parameters.npy")
testParameters = testParameters[:, 0:2]
testReflectances = np.load(dataFolder + "2015February2301:23AMNoisyRandomTestingreflectances1000000photons.npy")
testReflectances = mch.normalizeImageQuotient(testReflectances)
return trainingParameters, trainingReflectances, testParameters, testReflectances
def perfect(dataFolder):
trainingParameters = np.load(dataFolder + "2015February2401:43AMPerfectGridTraining10000parameters.npy")
trainingParameters = trainingParameters[:, 0:2] # only BVF and Vs for perfect data simulations
trainingReflectances = np.load(
dataFolder + "2015February2401:43AMPerfectGridTrainingreflectances1000000photons.npy"
)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
testParameters = np.load(dataFolder + "2015February2306:52PMUniformRandom10000parameters.npy")
testParameters = testParameters[:, 0:2] # only BVF and Vs for perfect data simulations
testReflectances = np.load(dataFolder + "2015February2306:52PMUniformRandomreflectances1000000photons.npy")
testReflectances = mch.normalizeImageQuotient(testReflectances)
return trainingParameters, trainingReflectances, testParameters, testReflectances
def realImage(dataFolder, imageToLoad):
trainingParameters = np.load(dataFolder + "2015February2807:13PMNoisyRandomTraining10000parameters.npy")
# estimate: BVF, Vs, d, SaO2:
trainingParameters = trainingParameters[:, 0:3]
trainingReflectances = np.load(
dataFolder + "2015February2807:13PMNoisyRandomTrainingreflectances1000000photons.npy"
)
# trainingReflectances = np.delete(trainingReflectances, [2, 7], axis=1)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
shape, image, trainsegmentation, testsegmentation = csvR.csvMultiSpectralImageReader2(dataFolder + imageToLoad)
# image = np.delete(image, [2, 7], axis=1)
image = mch.normalizeImageQuotient(image)
return trainingParameters, trainingReflectances, shape, image, trainsegmentation, testsegmentation
def realImage(dataFolder, imageToLoad):
trainingParameters = np.load(
dataFolder +
"2015February2807:13PMNoisyRandomTraining10000parameters.npy")
# estimate: BVF, Vs, d, SaO2:
trainingParameters = trainingParameters[:, 0:3]
trainingReflectances = np.load(
dataFolder +
"2015February2807:13PMNoisyRandomTrainingreflectances1000000photons.npy"
)
#trainingReflectances = np.delete(trainingReflectances, [2, 7], axis=1)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
shape, image, trainsegmentation, testsegmentation = csvR.csvMultiSpectralImageReader2(
dataFolder + imageToLoad)
#image = np.delete(image, [2, 7], axis=1)
image = mch.normalizeImageQuotient(image)
return trainingParameters, trainingReflectances, shape, image, trainsegmentation, testsegmentation
def noisy(dataFolder):
trainingParameters = np.load(
dataFolder +
"2015February2208:16PMNoisyRandomTraining10000parameters.npy")
trainingParameters = trainingParameters[:, 0:2]
trainingReflectances = np.load(
dataFolder +
"2015February2208:16PMNoisyRandomTrainingreflectances1000000photons.npy"
)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
testParameters = np.load(
dataFolder +
"2015February2301:23AMNoisyRandomTesting10000parameters.npy")
testParameters = testParameters[:, 0:2]
testReflectances = np.load(
dataFolder +
"2015February2301:23AMNoisyRandomTestingreflectances1000000photons.npy"
)
testReflectances = mch.normalizeImageQuotient(testReflectances)
return trainingParameters, trainingReflectances, testParameters, testReflectances
def perfect(dataFolder):
trainingParameters = np.load(
dataFolder +
"2015February2401:43AMPerfectGridTraining10000parameters.npy")
trainingParameters = trainingParameters[:, 0:
2] # only BVF and Vs for perfect data simulations
trainingReflectances = np.load(
dataFolder +
"2015February2401:43AMPerfectGridTrainingreflectances1000000photons.npy"
)
trainingReflectances = mch.normalizeImageQuotient(trainingReflectances)
testParameters = np.load(
dataFolder + "2015February2306:52PMUniformRandom10000parameters.npy")
testParameters = testParameters[:, 0:
2] # only BVF and Vs for perfect data simulations
testReflectances = np.load(
dataFolder +
"2015February2306:52PMUniformRandomreflectances1000000photons.npy")
testReflectances = mch.normalizeImageQuotient(testReflectances)
return trainingParameters, trainingReflectances, testParameters, testReflectances
def estimateParametersRealImage(trainingParameters, trainingReflectances,
shape, image, trainsegmentation,
testsegmentation, activateDA):
sourceReflectancesDA = image[np.nonzero(trainsegmentation)[0], :]
# choose m reflectances for training DA
m = trainingReflectances.shape[0]
sourceReflectancesDA = np.matrix(random.sample(sourceReflectancesDA, m))
#%% 2. determine domain adaptation weights
trainingWeights = np.ones(trainingReflectances.shape[0])
if (activateDA):
trainingWeights = calculateWeights(trainingReflectances,
sourceReflectancesDA)
#%% 3. train forest
rf = randomForest(trainingParameters, trainingReflectances,
trainingWeights)
#%% 4. estimate the parameters for the image
print "starting to estimate the tissue parameters"
start = time.time()
estimatedParameters = rf.predict(image)
# set to zero if not in segmentation mask
#estimatedParameters[np.where(0 == testsegmentation), :] = 0
end = time.time()
print "time necessary to estimate parameters for image [s]: " + str(
(end - start))
#%% save the parametric images TODO delete after everything works
# import Image
#
# for i in np.arange(0,estimatedParameters.shape[1]):
# parameterImage_i = np.reshape(estimatedParameters[:,i], shape)
# im = Image.fromarray(parameterImage_i)
# im.save("data/output/" + "parameterImage_" + str(i) + ".tiff")
#%% 6. evaluate data
# for this, create monte carlo simulation for each
# parameter estimate. The resulted reflectance estimate can then be compared to
# the measured reflectance.
from setup import systemPaths
from setup import simulation
import helper.monteCarloHelper as mch
infileString, outfolderMC, outfolderRS, gpumcmlDirectory, gpumcmlExecutable = systemPaths.initPaths(
)
infile = open(infileString)
BVFs, Vss, ds, SaO2s, rs, nrSamples, photons, wavelengths, FWHM, eHbO2, eHb, nrSimulations = simulation.noisy(
)
# the estimated parameters within the segmentation
estimatedParametersOnlySegmented = estimatedParameters[
np.nonzero(testsegmentation)[0], :]
# the image reflectances from which these parameters where estimated
inputReflectancesOnlySegmented = image[np.nonzero(testsegmentation)[0], :]
# index vector for selecting n samples from this data
indices = np.arange(0, estimatedParametersOnlySegmented.shape[0], 1)
# choose n
n = 20
nSamples = random.sample(indices, n)
estimatedParametersOnlySegmented = estimatedParametersOnlySegmented[
nSamples]
inputReflectancesOnlySegmented = inputReflectancesOnlySegmented[nSamples]
# placeholder for the reflectance computed from MC with the estimated parameters
reflectancesFromEstimatedParameters = np.zeros(
(inputReflectancesOnlySegmented.shape[0],
inputReflectancesOnlySegmented.shape[1] + 1))
#wavelengths = np.delete(wavelengths, [2, 7])
for i, (BVF, Vs, d) in enumerate(estimatedParametersOnlySegmented):
print('starting simulation ' + str(i) + ' of ' +
str(estimatedParametersOnlySegmented.shape[0]))
for j, wavelength in enumerate(wavelengths):
reflectanceValue = mch.runOneSimulation(
wavelength,
eHbO2,
eHb,
infile,
outfolderMC,
gpumcmlDirectory,
gpumcmlExecutable,
BVF,
Vs,
d,
# np.mean(rs), SaO2,
# submucosa_BVF=sm_BVF, submucosa_Vs=sm_Vs, submucosa_SaO2=SaO2,
Fwhm=FWHM,
nrPhotons=photons)
#print((BVF, Vs, d, wavelength))
reflectancesFromEstimatedParameters[i, j] = reflectanceValue
# correct these reflectances by image quotient
reflectancesFromEstimatedParameters = mch.normalizeImageQuotient(
reflectancesFromEstimatedParameters)
wavelengths = mch.removeIqWavelength(wavelengths)
#%% plot data for nicer inspection
from sklearn.metrics import r2_score
r2Score = r2_score(reflectancesFromEstimatedParameters.T,
inputReflectancesOnlySegmented.T)
print("r2Score for random forest estimatation of:", str(r2Score))
#%% sort by wavelength:
for plot_i in range(n):
sortedIndices = sorted(range(len(wavelengths)),
key=lambda k: wavelengths[k])
plt.figure()
plt.plot(wavelengths[sortedIndices],
reflectancesFromEstimatedParameters[plot_i,
sortedIndices], 'g-o')
plt.plot(wavelengths[sortedIndices],
inputReflectancesOnlySegmented[plot_i, sortedIndices], 'b-o')
print(
str(
r2_score(reflectancesFromEstimatedParameters[plot_i, :],
inputReflectancesOnlySegmented[plot_i, :])))
plt.legend(["estimated", "measurement"])
plt.xlabel("wavelength [m]")
plt.ylabel("normalized reflectance")
plt.savefig("data/output/example_fit_" + str(plot_i) + '.png')
return estimatedParameters, r2Score, reflectancesFromEstimatedParameters, inputReflectancesOnlySegmented
def estimateParametersRealImage(trainingParameters, trainingReflectances, shape, image, trainsegmentation, testsegmentation, activateDA):
sourceReflectancesDA = image[np.nonzero(trainsegmentation)[0], :]
# choose m reflectances for training DA
m = trainingReflectances.shape[0]
sourceReflectancesDA = np.matrix(random.sample(sourceReflectancesDA, m))
#%% 2. determine domain adaptation weights
trainingWeights = np.ones(trainingReflectances.shape[0])
if (activateDA):
trainingWeights = calculateWeights(trainingReflectances, sourceReflectancesDA)
#%% 3. train forest
rf = randomForest(trainingParameters, trainingReflectances, trainingWeights)
#%% 4. estimate the parameters for the image
print "starting to estimate the tissue parameters"
start = time.time()
estimatedParameters = rf.predict(image)
# set to zero if not in segmentation mask
#estimatedParameters[np.where(0 == testsegmentation), :] = 0
end = time.time()
print "time necessary to estimate parameters for image [s]: " + str((end - start))
#%% save the parametric images TODO delete after everything works
# import Image
#
# for i in np.arange(0,estimatedParameters.shape[1]):
# parameterImage_i = np.reshape(estimatedParameters[:,i], shape)
# im = Image.fromarray(parameterImage_i)
# im.save("data/output/" + "parameterImage_" + str(i) + ".tiff")
#%% 6. evaluate data
# for this, create monte carlo simulation for each
# parameter estimate. The resulted reflectance estimate can then be compared to
# the measured reflectance.
from setup import systemPaths
from setup import simulation
import helper.monteCarloHelper as mch
infileString, outfolderMC, outfolderRS, gpumcmlDirectory, gpumcmlExecutable = systemPaths.initPaths()
infile = open(infileString)
BVFs, Vss, ds, SaO2s, rs, nrSamples, photons, wavelengths, FWHM, eHbO2, eHb, nrSimulations = simulation.noisy()
# the estimated parameters within the segmentation
estimatedParametersOnlySegmented = estimatedParameters[np.nonzero(testsegmentation)[0], :]
# the image reflectances from which these parameters where estimated
inputReflectancesOnlySegmented = image[np.nonzero(testsegmentation)[0], :]
# index vector for selecting n samples from this data
indices = np.arange(0, estimatedParametersOnlySegmented.shape[0], 1)
# choose n
n = 20
nSamples = random.sample(indices, n)
estimatedParametersOnlySegmented = estimatedParametersOnlySegmented[nSamples]
inputReflectancesOnlySegmented = inputReflectancesOnlySegmented[nSamples]
# placeholder for the reflectance computed from MC with the estimated parameters
reflectancesFromEstimatedParameters = np.zeros((inputReflectancesOnlySegmented.shape[0], inputReflectancesOnlySegmented.shape[1]+1))
#wavelengths = np.delete(wavelengths, [2, 7])
for i, (BVF, Vs, d) in enumerate(estimatedParametersOnlySegmented):
print('starting simulation ' + str(i) + ' of ' + str(estimatedParametersOnlySegmented.shape[0]))
for j, wavelength in enumerate(wavelengths):
reflectanceValue = mch.runOneSimulation(
wavelength, eHbO2, eHb,
infile, outfolderMC, gpumcmlDirectory, gpumcmlExecutable,
BVF, Vs, d,
# np.mean(rs), SaO2,
# submucosa_BVF=sm_BVF, submucosa_Vs=sm_Vs, submucosa_SaO2=SaO2,
Fwhm = FWHM, nrPhotons=photons)
#print((BVF, Vs, d, wavelength))
reflectancesFromEstimatedParameters[i, j] = reflectanceValue
# correct these reflectances by image quotient
reflectancesFromEstimatedParameters = mch.normalizeImageQuotient(reflectancesFromEstimatedParameters)
wavelengths = mch.removeIqWavelength(wavelengths)
#%% plot data for nicer inspection
from sklearn.metrics import r2_score
r2Score = r2_score(reflectancesFromEstimatedParameters.T, inputReflectancesOnlySegmented.T)
print("r2Score for random forest estimatation of:", str(r2Score))
#%% sort by wavelength:
for plot_i in range(n):
sortedIndices = sorted(range(len(wavelengths)), key=lambda k: wavelengths[k])
plt.figure()
plt.plot(wavelengths[sortedIndices], reflectancesFromEstimatedParameters[plot_i,sortedIndices], 'g-o')
plt.plot(wavelengths[sortedIndices], inputReflectancesOnlySegmented[plot_i,sortedIndices], 'b-o')
print(str(r2_score(reflectancesFromEstimatedParameters[plot_i, :], inputReflectancesOnlySegmented[plot_i, :])))
plt.legend(["estimated", "measurement"])
plt.xlabel("wavelength [m]")
plt.ylabel("normalized reflectance")
plt.savefig("data/output/example_fit_" + str(plot_i) + '.png')
return estimatedParameters, r2Score, reflectancesFromEstimatedParameters, inputReflectancesOnlySegmented