def predictContours(self):
'''
Method for pushing images loaded by _read_all_dicom_images through full learned network
to predict the resulting contour
STEPS: 1) Feedforward through CNN, followed by through SA, followed by calling AC model, giving final result
:return:
'''
#images are slice * time * height * width
self.predROIs = np.array([CNNpred(inputimages = self.images[s,:], batch_size=1,
fine_tuned_params_path = '/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/fine_tune_paramsXnew.pickle')
for s in range(0, len(self.slices))])
self.imagesROIs = np.array([crop_ROI(images=self.images[s,:], roi=self.predROIs[s,:],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.slices))])
self.predSAContours = np.array([SApred(self.imagesROIs[s,:],
trained_SA_path ='/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SA_XPHmodel')
for s in range(0, len(self.slices))])
self.predACContours = np.array([[AC.evolve_contour(lv = self.predSAContours[s,t], roi=self.imagesROIs[s,t],alpha3=0)
for t in range(0, len(self.time))] for s in range(0, len(self.slices))])
def predictContours(self):
'''
Method for pushing images loaded by _read_all_dicom_images through full learned network
to predict the resulting contour
STEPS: 1) Feedforward through CNN, followed by through SA, followed by calling AC model, giving final result
:return:
'''
#images are slice * time * height * width
self.predROIs = np.array([
CNNpred(
inputimages=self.images[s, :],
batch_size=1,
fine_tuned_params_path=
'/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/fine_tune_paramsXnew.pickle'
) for s in range(0, len(self.slices))
])
self.imagesROIs = np.array([
crop_ROI(images=self.images[s, :],
roi=self.predROIs[s, :],
roi_dim=(100, 100),
newsize=(64, 64)) for s in range(0, len(self.slices))
])
self.predSAContours = np.array([
SApred(
self.imagesROIs[s, :],
trained_SA_path=
'/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SA_XPHmodel'
) for s in range(0, len(self.slices))
])
self.predACContours = np.array([[
AC.evolve_contour(lv=self.predSAContours[s, t],
roi=self.imagesROIs[s, t],
alpha3=0) for t in range(0, len(self.time))
] for s in range(0, len(self.slices))])
def predictContours(self):
'''
Method for pushing images loaded by _read_all_dicom_images through full learned network
to predict the resulting contour
STEPS: 1) Feedforward through CNN, followed by through SA, followed by calling AC model, giving final result
:return:
'''
#images are slice * time * height * width
#############################################
# PREDICT ROI MASKS
# IMPORTANT: Change fine_tuned_params_path file to fine_tune_paramsX_big.pickle and .._small.pickle respectively
# self.big_images and self.small_images are both lists of lists of arrays
b = len(self.big_images)
s = len(self.small_images)
# CNN should return list of arrays
self.pred_big_ROIs = [CNNpred(inputimages = self.big_images[s], batch_size=1,
fine_tuned_params_path = '/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/fine_tune_paramsXlarge.pickle')
for s in range(0, len(self.big_slices))]
self.pred_small_ROIs = [CNNpred(inputimages = self.small_images[s], batch_size=1,
fine_tuned_params_path = '/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/fine_tune_paramsXsmall.pickle')
for s in range(0, len(self.small_slices))]
"""
self.predROIs = np.array([CNNpred(inputimages = self.images[s,:], batch_size=1,
fine_tuned_params_path = '/Users/mh/AML/DataScienceBowl/data/fine_tune_paramsXnew.pickle')
for s in range(0, len(self.slices))])
"""
#############################################
# CROP IMAGES TO ROIs
# crop ROI should return list of arrays
self.imagesROIs_big = [crop_ROI(images=self.big_images[s], roi=self.pred_big_ROIs[s],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.big_slices))]
self.imagesROIs_small = [crop_ROI(images=self.small_images[s], roi=self.pred_small_ROIs[s],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.small_slices))]
"""
self.imagesROIs = np.array([crop_ROI(images=self.images[s,:], roi=self.predROIs[s,:],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.slices))])
"""
#############################################
# PREDICT CONTOUR USING SA
# IMPORTANT: Change trained_SA_path to ../SA_Xmodel_big and ../SA_Xmodel_small, respectively
self.predSAbigContours = np.array([SApred(self.imagesROIs_big[s],
trained_SA_path ='/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/SA_RLUmodel_large')
for s in range(0, len(self.big_slices))])
self.predSAsmallContours = np.array([SApred(self.imagesROIs_small[s],
trained_SA_path ='/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/SA_RLUmodel_small')
for s in range(0, len(self.small_slices))])
"""
self.predSAContours = np.array([SApred(self.imagesROIs[s,:],
trained_SA_path ='/Users/mh/AML/DataScienceBowl/data/SA_Xmodel')
for s in range(0, len(self.slices))])
"""
#############################################
# ACTIVE CONTOUR
self.predACContours_big = np.array([[AC.evolve_contour(lv = self.predSAbigContours[s][t], roi=self.imagesROIs_big[s][t]
, alpha1=2, alpha2=1.5, alpha3=0.002)
for t in range(0, len(self.time))] for s in range(0, len(self.big_slices))])
self.predACContours_small = np.array([[AC.evolve_contour(lv = self.predSAsmallContours[s][t], roi=self.imagesROIs_small[s][t]
, alpha1=2, alpha2=1.5, alpha3=0.007)
for t in range(0, len(self.time))] for s in range(0, len(self.small_slices))])
"""
# original training image
train_large = np.load(LOCALDATAPATH + 'SBXtrainImage256_large')
train_small = np.load(LOCALDATAPATH + 'SBXtrainImage256_small')
# original training binary contour
mask_large = np.load(LOCALDATAPATH + 'SBXtrainMask256_large')
mask_small = np.load(LOCALDATAPATH + 'SBXtrainMask256_small')
# predictions from SA based on the above data
preds_Large = np.load(LOCALDATAPATH + 'SA_predictions_large')
preds_Small = np.load(LOCALDATAPATH + 'SA_predictions_small')
# crop original image and contour data to match the region of the predictions
train_roi_large = crop_ROI(images=train_large,
roi=roi_large,
roi_dim=(100, 100),
newsize=(64, 64))
train_roi_small = crop_ROI(images=train_small,
roi=roi_small,
roi_dim=(100, 100),
newsize=(64, 64))
mask_roi_large = crop_ROI(images=mask_large,
roi=roi_large,
roi_dim=(100, 100),
newsize=(64, 64))
mask_roi_small = crop_ROI(images=mask_small,
roi=roi_small,
roi_dim=(100, 100),
newsize=(64, 64))
masks = np.array(mask_predictions)
return masks
if __name__ == "__main__":
# load required data
roi = np.load('/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SBXtrainBinaryMask32')
train = np.load('/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SBXtrainImage256')
mask = np.load('/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SBXtrainMask256')
# crop data based on ROI
train_roi =crop_ROI(images=train, roi=roi, roi_dim=(100,100), newsize=(64, 64))
mask_roi =crop_ROI(images= mask, roi=roi, roi_dim=(100,100), newsize=(64, 64))
# call predict function
mask_predictions = predict_sa(train_roi, trained_SA_path = '/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SA_RLUmodel')
# store predictions
mask_predictions.dump('/Users/Peadar/Documents/KagglePythonProjects/AML/DataScienceBowl/data/SA_predictions')
# methods for plotting scatter of results
for i in range(0,100):
f, axarr = plt.subplots(3, 2)
axarr[0, 1].imshow(train[i,:,:])
axarr[0, 1].set_title('Original Image')
def predictContours(self):
'''
Method for pushing images loaded by _read_all_dicom_images through full learned network
to predict the resulting contour
STEPS: 1) Feedforward through CNN, followed by through SA, followed by calling AC model, giving final result
:return:
'''
#images are slice * time * height * width
#############################################
# PREDICT ROI MASKS
# IMPORTANT: Change fine_tuned_params_path file to fine_tune_paramsX_big.pickle and .._small.pickle respectively
# self.big_images and self.small_images are both lists of lists of arrays
b = len(self.big_images)
s = len(self.small_images)
# CNN should return list of arrays
self.pred_big_ROIs = [CNNpred(inputimages = self.big_images[s], batch_size=1,
fine_tuned_params_path = '/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/fine_tune_paramsXlarge.pickle')
for s in range(0, len(self.big_slices))]
self.pred_small_ROIs = [CNNpred(inputimages = self.small_images[s], batch_size=1,
fine_tuned_params_path = '/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/fine_tune_paramsXsmall.pickle')
for s in range(0, len(self.small_slices))]
"""
self.predROIs = np.array([CNNpred(inputimages = self.images[s,:], batch_size=1,
fine_tuned_params_path = '/Users/mh/AML/DataScienceBowl/data/fine_tune_paramsXnew.pickle')
for s in range(0, len(self.slices))])
"""
#############################################
# CROP IMAGES TO ROIs
# crop ROI should return list of arrays
self.imagesROIs_big = [crop_ROI(images=self.big_images[s], roi=self.pred_big_ROIs[s],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.big_slices))]
self.imagesROIs_small = [crop_ROI(images=self.small_images[s], roi=self.pred_small_ROIs[s],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.small_slices))]
"""
self.imagesROIs = np.array([crop_ROI(images=self.images[s,:], roi=self.predROIs[s,:],
roi_dim=(100,100), newsize=(64, 64))
for s in range(0, len(self.slices))])
"""
#############################################
# PREDICT CONTOUR USING SA
# IMPORTANT: Change trained_SA_path to ../SA_Xmodel_big and ../SA_Xmodel_small, respectively
self.predSAbigContours = np.array([SApred(self.imagesROIs_big[s],
trained_SA_path ='/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/SA_RLUmodel_large')
for s in range(0, len(self.big_slices))])
self.predSAsmallContours = np.array([SApred(self.imagesROIs_small[s],
trained_SA_path ='/home/odyss/Desktop/dsb/AML/DataScienceBowl/data/SA_RLUmodel_small')
for s in range(0, len(self.small_slices))])
"""
self.predSAContours = np.array([SApred(self.imagesROIs[s,:],
trained_SA_path ='/Users/mh/AML/DataScienceBowl/data/SA_Xmodel')
for s in range(0, len(self.slices))])
"""
#############################################
# ACTIVE CONTOUR
self.predACContours_big = np.array([[AC.evolve_contour(lv = self.predSAbigContours[s][t], roi=self.imagesROIs_big[s][t]
, alpha1=2, alpha2=1.5, alpha3=0.002)
for t in range(0, len(self.time))] for s in range(0, len(self.big_slices))])
self.predACContours_small = np.array([[AC.evolve_contour(lv = self.predSAsmallContours[s][t], roi=self.imagesROIs_small[s][t]
, alpha1=2, alpha2=1.5, alpha3=0.007)
for t in range(0, len(self.time))] for s in range(0, len(self.small_slices))])
"""
roi_small = np.load(LOCALDATAPATH + 'SBXtrainBinaryMask32_small')
# original training image
train_large = np.load(LOCALDATAPATH + 'SBXtrainImage256_large')
train_small = np.load(LOCALDATAPATH + 'SBXtrainImage256_small')
# original training binary contour
mask_large = np.load(LOCALDATAPATH + 'SBXtrainMask256_large')
mask_small = np.load(LOCALDATAPATH + 'SBXtrainMask256_small')
# predictions from SA based on the above data
preds_Large = np.load(LOCALDATAPATH + 'SA_predictions_large')
preds_Small = np.load(LOCALDATAPATH + 'SA_predictions_small')
# crop original image and contour data to match the region of the predictions
train_roi_large =crop_ROI(images=train_large, roi=roi_large, roi_dim=(100,100), newsize=(64, 64))
train_roi_small =crop_ROI(images=train_small, roi=roi_small, roi_dim=(100,100), newsize=(64, 64))
mask_roi_large =crop_ROI(images=mask_large, roi=roi_large, roi_dim=(100,100), newsize=(64, 64))
mask_roi_small =crop_ROI(images=mask_small, roi=roi_small, roi_dim=(100,100), newsize=(64, 64))
show_preds_large = [train_roi_large[i,:,:] + preds_Large[i,:,:] for i in range(np.shape(preds_Large)[0])]
show_preds_small = [train_roi_small[i,:,:] + preds_Small[i,:,:] for i in range(np.shape(preds_Small)[0])]
#pdb.set_trace()
"""
Trial parameter ranges: alpha1{1, 1.5, 2}, alpha 2{1.5,2,2.5}, alpha 3 = {0, ..., 0.01} steps 0.001
"""
