def run(img_folder,
img_extension='dcm',
img_size=[288, 224],
img_scale=4095,
multi_view=False,
do_featurewise_norm=True,
featurewise_mean=398.5,
featurewise_std=627.8,
batch_size=16,
samples_per_epoch=160,
nb_epoch=20,
balance_classes=.0,
all_neg_skip=0.,
pos_cls_weight=1.0,
nb_init_filter=64,
init_filter_size=7,
init_conv_stride=2,
pool_size=3,
pool_stride=2,
weight_decay=.0001,
alpha=1.,
l1_ratio=.5,
inp_dropout=.0,
hidden_dropout=.0,
init_lr=.01,
val_size=.2,
lr_patience=5,
es_patience=10,
resume_from=None,
net='resnet50',
load_val_ram=False,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
best_model='./modelState/dm_resnet_best_model.h5',
final_model="NOSAVE"):
'''Run ResNet training on mammograms using an exam or image list
Args:
featurewise_mean, featurewise_std ([float]): they are estimated from
1152 x 896 images. Using different sized images give very close
results. For png, mean=7772, std=12187.
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Setup training and validation data.
# Load image or exam lists and split them into train and val sets.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
if multi_view:
exam_list = meta_man.get_flatten_exam_list()
exam_train, exam_val = train_test_split(
exam_list,
test_size=val_size,
random_state=random_seed,
stratify=meta_man.exam_labs(exam_list))
val_size_ = len(exam_val) * 2 # L and R.
else:
img_list, lab_list = meta_man.get_flatten_img_list()
img_train, img_val, lab_train, lab_val = train_test_split(
img_list,
lab_list,
test_size=val_size,
random_state=random_seed,
stratify=lab_list)
val_size_ = len(img_val)
# Create image generator.
img_gen = DMImageDataGenerator(horizontal_flip=True, vertical_flip=True)
if do_featurewise_norm:
img_gen.featurewise_center = True
img_gen.featurewise_std_normalization = True
img_gen.mean = featurewise_mean
img_gen.std = featurewise_std
else:
img_gen.samplewise_center = True
img_gen.samplewise_std_normalization = True
if multi_view:
train_generator = img_gen.flow_from_exam_list(
exam_train,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
if load_val_ram:
val_generator = img_gen.flow_from_exam_list(
exam_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=val_size_,
validation_mode=True,
class_mode='binary')
else:
val_generator = img_gen.flow_from_exam_list(
exam_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
else:
train_generator = img_gen.flow_from_img_list(
img_train,
lab_train,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
if load_val_ram:
val_generator = img_gen.flow_from_img_list(
img_val,
lab_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=val_size_,
validation_mode=True,
class_mode='binary')
else:
val_generator = img_gen.flow_from_img_list(
img_val,
lab_val,
target_size=(img_size[0], img_size[1]),
target_scale=img_scale,
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
# Load validation set into RAM.
if load_val_ram:
validation_set = next(val_generator)
if not multi_view and len(validation_set[0]) != val_size_:
raise Exception
elif len(validation_set[0][0]) != val_size_ \
or len(validation_set[0][1]) != val_size_:
raise Exception
# Create model.
if resume_from is not None:
model = load_model(resume_from,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
else:
if multi_view:
builder = MultiViewResNetBuilder
else:
builder = ResNetBuilder
if net == 'resnet18':
model = builder.build_resnet_18(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'resnet34':
model = builder.build_resnet_34(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'resnet50':
model = builder.build_resnet_50(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet14':
model = builder.build_dm_resnet_14(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet47rb5':
model = builder.build_dm_resnet_47rb5(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet56rb6':
model = builder.build_dm_resnet_56rb6(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'dmresnet65rb7':
model = builder.build_dm_resnet_65rb7(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'resnet101':
model = builder.build_resnet_101(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
elif net == 'resnet152':
model = builder.build_resnet_152(
(1, img_size[0], img_size[1]), 1, nb_init_filter,
init_filter_size, init_conv_stride, pool_size, pool_stride,
weight_decay, alpha, l1_ratio, inp_dropout, hidden_dropout)
if gpu_count > 1:
model = make_parallel(model, gpu_count)
# Model training.
sgd = SGD(lr=init_lr, momentum=0.9, decay=0.0, nesterov=True)
model.compile(optimizer=sgd,
loss='binary_crossentropy',
metrics=[DMMetrics.sensitivity, DMMetrics.specificity])
reduce_lr = ReduceLROnPlateau(monitor='val_loss',
factor=0.1,
patience=lr_patience,
verbose=1)
early_stopping = EarlyStopping(monitor='val_loss',
patience=es_patience,
verbose=1)
if load_val_ram:
auc_checkpointer = DMAucModelCheckpoint(best_model,
validation_set,
batch_size=batch_size)
else:
auc_checkpointer = DMAucModelCheckpoint(best_model,
val_generator,
nb_test_samples=val_size_)
# checkpointer = ModelCheckpoint(
# best_model, monitor='val_loss', verbose=1, save_best_only=True)
hist = model.fit_generator(
train_generator,
samples_per_epoch=samples_per_epoch,
nb_epoch=nb_epoch,
class_weight={
0: 1.0,
1: pos_cls_weight
},
validation_data=validation_set if load_val_ram else val_generator,
nb_val_samples=val_size_,
callbacks=[reduce_lr, early_stopping, auc_checkpointer],
nb_worker=nb_worker,
pickle_safe=True, # turn on pickle_safe to avoid a strange error.
verbose=2)
# Training report.
min_loss_locs, = np.where(
hist.history['val_loss'] == min(hist.history['val_loss']))
best_val_loss = hist.history['val_loss'][min_loss_locs[0]]
best_val_sensitivity = hist.history['val_sensitivity'][min_loss_locs[0]]
best_val_specificity = hist.history['val_specificity'][min_loss_locs[0]]
print "\n==== Training summary ===="
print "Minimum val loss achieved at epoch:", min_loss_locs[0] + 1
print "Best val loss:", best_val_loss
print "Best val sensitivity:", best_val_sensitivity
print "Best val specificity:", best_val_specificity
if final_model != "NOSAVE":
model.save(final_model)
return hist
def run(img_folder, dl_state, img_extension='dcm',
img_height=1024, img_scale=4095, val_size=.2, neg_vs_pos_ratio=10.,
do_featurewise_norm=True, featurewise_mean=873.6, featurewise_std=739.3,
img_per_batch=2, roi_per_img=32, roi_size=(256, 256),
low_int_threshold=.05, blob_min_area=3,
blob_min_int=.5, blob_max_int=.85, blob_th_step=10,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
train_out='./modelState/meta_prob_train.pkl',
test_out='./modelState/meta_prob_test.pkl'):
'''Calculate bag of deep visual words count matrix for all breasts
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_train, subj_test, labs_train, labs_test = train_test_split(
subj_list, subj_labs, test_size=val_size, stratify=subj_labs,
random_state=random_seed)
if neg_vs_pos_ratio is not None:
def subset_subj(subj, labs):
subj = np.array(subj)
labs = np.array(labs)
pos_idx = np.where(labs==1)[0]
neg_idx = np.where(labs==0)[0]
nb_neg_desired = int(len(pos_idx)*neg_vs_pos_ratio)
if nb_neg_desired >= len(neg_idx):
return subj.tolist()
else:
neg_chosen = rng.choice(neg_idx, nb_neg_desired, replace=False)
subset_idx = np.concatenate([pos_idx, neg_chosen])
return subj[subset_idx].tolist()
subj_train = subset_subj(subj_train, labs_train)
subj_test = subset_subj(subj_test, labs_test)
# Create image generator for ROIs for representation extraction.
print "Create an image generator for ROIs"; sys.stdout.flush()
if do_featurewise_norm:
imgen = DMImageDataGenerator(
featurewise_center=True,
featurewise_std_normalization=True)
imgen.mean = featurewise_mean
imgen.std = featurewise_std
else:
imgen = DMImageDataGenerator(
samplewise_center=True,
samplewise_std_normalization=True)
# Load DL model.
print "Load DL classification model:", dl_state; sys.stdout.flush()
dl_model = load_model(
dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
}
)
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model = make_parallel(dl_model, gpu_count)
# Read exam lists.
exam_train = meta_man.get_flatten_exam_list(
subj_train, flatten_img_list=True)
exam_test = meta_man.get_flatten_exam_list(
subj_test, flatten_img_list=True)
exam_labs_train = np.array(meta_man.exam_labs(exam_train))
exam_labs_test = np.array(meta_man.exam_labs(exam_test))
nb_pos_exams_train = (exam_labs_train==1).sum()
nb_neg_exams_train = (exam_labs_train==0).sum()
nb_pos_exams_test = (exam_labs_test==1).sum()
nb_neg_exams_test = (exam_labs_test==0).sum()
print "Train set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_train, nb_neg_exams_train)
print "Test set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_test, nb_neg_exams_test)
# Make predictions for exam lists.
print "Predicting for train exam list"; sys.stdout.flush()
meta_prob_train = get_exam_pred(
exam_train, roi_per_img, imgen,
target_height=img_height, target_scale=img_scale,
img_per_batch=img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=dl_model)
print "Length of train prediction list:", len(meta_prob_train)
sys.stdout.flush()
print "Predicting for test exam list"; sys.stdout.flush()
meta_prob_test = get_exam_pred(
exam_test, roi_per_img, imgen,
target_height=img_height, target_scale=img_scale,
img_per_batch=img_per_batch, roi_size=roi_size,
low_int_threshold=low_int_threshold, blob_min_area=blob_min_area,
blob_min_int=blob_min_int, blob_max_int=blob_max_int,
blob_th_step=blob_th_step, seed=random_seed,
dl_model=dl_model)
print "Length of test prediction list:", len(meta_prob_test)
sys.stdout.flush()
pickle.dump(meta_prob_train, open(train_out, 'w'))
pickle.dump(meta_prob_test, open(test_out, 'w'))
print "Done."
def run(img_folder,
img_extension='png',
img_size=[288, 224],
multi_view=False,
do_featurewise_norm=True,
featurewise_mean=7772.,
featurewise_std=12187.,
batch_size=16,
samples_per_epoch=160,
nb_epoch=20,
val_size=.2,
balance_classes=0.,
all_neg_skip=False,
pos_cls_weight=1.0,
alpha=1.,
l1_ratio=.5,
init_lr=.01,
lr_patience=2,
es_patience=4,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
dl_state='./modelState/resnet50_288_best_model.h5',
best_model='./modelState/enet_288_best_model.h5',
final_model="NOSAVE"):
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
nb_worker = int(os.getenv('NUM_CPU_CORES', 4))
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Setup training and validation data.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
if multi_view:
exam_list = meta_man.get_flatten_exam_list()
exam_train, exam_val = train_test_split(
exam_list,
test_size=val_size,
random_state=random_seed,
stratify=meta_man.exam_labs(exam_list))
val_size_ = len(exam_val) * 2 # L and R.
else:
img_list, lab_list = meta_man.get_flatten_img_list()
img_train, img_val, lab_train, lab_val = train_test_split(
img_list,
lab_list,
test_size=val_size,
random_state=random_seed,
stratify=lab_list)
val_size_ = len(img_val)
img_gen = DMImageDataGenerator(horizontal_flip=True, vertical_flip=True)
if do_featurewise_norm:
img_gen.featurewise_center = True
img_gen.featurewise_std_normalization = True
img_gen.mean = featurewise_mean
img_gen.std = featurewise_std
else:
img_gen.samplewise_center = True
img_gen.samplewise_std_normalization = True
if multi_view:
train_generator = img_gen.flow_from_exam_list(
exam_train,
target_size=(img_size[0], img_size[1]),
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
val_generator = img_gen.flow_from_exam_list(exam_val,
target_size=(img_size[0],
img_size[1]),
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
else:
train_generator = img_gen.flow_from_img_list(
img_train,
lab_train,
target_size=(img_size[0], img_size[1]),
batch_size=batch_size,
balance_classes=balance_classes,
all_neg_skip=all_neg_skip,
shuffle=True,
seed=random_seed,
class_mode='binary')
val_generator = img_gen.flow_from_img_list(img_val,
lab_val,
target_size=(img_size[0],
img_size[1]),
batch_size=batch_size,
validation_mode=True,
class_mode='binary')
# Deep learning model.
dl_model = load_model(dl_state,
custom_objects={
'sensitivity': DMMetrics.sensitivity,
'specificity': DMMetrics.specificity
})
# Dummy compilation to turn off the "uncompiled" error when model was run on multi-GPUs.
# dl_model.compile(optimizer='sgd', loss='binary_crossentropy')
reprlayer_model = Model(input=dl_model.input,
output=dl_model.get_layer(index=-2).output)
if gpu_count > 1:
reprlayer_model = make_parallel(reprlayer_model, gpu_count)
# Setup test data in RAM.
X_test, y_test = dlrepr_generator(reprlayer_model, val_generator,
val_size_)
# import pdb; pdb.set_trace()
# Evaluat DL model on the test data.
val_generator.reset()
dl_test_pred = dl_model.predict_generator(val_generator,
val_samples=val_size_,
nb_worker=1,
pickle_safe=False)
# Set nb_worker to >1 can cause:
# either inconsistent result when pickle_safe is False,
# or broadcasting error when pickle_safe is True.
# This seems to be a Keras bug!!
# Further note: the broadcasting error may only happen when val_size_
# is not divisible by batch_size.
try:
dl_auc = roc_auc_score(y_test, dl_test_pred)
dl_loss = log_loss(y_test, dl_test_pred)
except ValueError:
dl_auc = 0.
dl_loss = np.inf
print "\nAUROC by the DL model: %.4f, loss: %.4f" % (dl_auc, dl_loss)
# import pdb; pdb.set_trace()
# Elastic net training.
target_classes = np.array([0, 1])
sgd_clf = SGDClassifier(loss='log',
penalty='elasticnet',
alpha=alpha,
l1_ratio=l1_ratio,
verbose=0,
n_jobs=nb_worker,
learning_rate='constant',
eta0=init_lr,
random_state=random_seed,
class_weight={
0: 1.0,
1: pos_cls_weight
})
curr_lr = init_lr
best_epoch = 0
best_auc = 0.
min_loss = np.inf
min_loss_epoch = 0
for epoch in xrange(nb_epoch):
samples_seen = 0
X_list = []
y_list = []
epoch_start = time.time()
while samples_seen < samples_per_epoch:
X, y = next(train_generator)
X_repr = reprlayer_model.predict_on_batch(X)
sgd_clf.partial_fit(X_repr, y, classes=target_classes)
samples_seen += len(y)
X_list.append(X_repr)
y_list.append(y)
# The training X, y are expected to change for each epoch due to
# image random sampling and class balancing.
X_train_epo = np.concatenate(X_list)
y_train_epo = np.concatenate(y_list)
# End of epoch summary.
pred_prob = sgd_clf.predict_proba(X_test)[:, 1]
train_prob = sgd_clf.predict_proba(X_train_epo)[:, 1]
try:
auc = roc_auc_score(y_test, pred_prob)
crossentropy_loss = log_loss(y_test, pred_prob)
except ValueError:
auc = 0.
crossentropy_loss = np.inf
try:
train_loss = log_loss(y_train_epo, train_prob)
except ValueError:
train_loss = np.inf
wei_sparseness = np.mean(sgd_clf.coef_ == 0)
epoch_span = time.time() - epoch_start
print ("%ds - Epoch=%d, auc=%.4f, train_loss=%.4f, test_loss=%.4f, "
"weight sparsity=%.4f") % \
(epoch_span, epoch + 1, auc, train_loss, crossentropy_loss,
wei_sparseness)
# Model checkpoint, reducing learning rate and early stopping.
if auc > best_auc:
best_epoch = epoch + 1
best_auc = auc
if best_model != "NOSAVE":
with open(best_model, 'w') as best_state:
pickle.dump(sgd_clf, best_state)
if crossentropy_loss < min_loss:
min_loss = crossentropy_loss
min_loss_epoch = epoch + 1
else:
if epoch + 1 - min_loss_epoch >= es_patience:
print 'Early stopping criterion has reached. Stop training.'
break
if epoch + 1 - min_loss_epoch >= lr_patience:
curr_lr *= .1
sgd_clf.set_params(eta0=curr_lr)
print "Reducing learning rate to: %s" % (curr_lr)
# End of training summary
print ">>> Found best AUROC: %.4f at epoch: %d, saved to: %s <<<" % \
(best_auc, best_epoch, best_model)
print ">>> Found best val loss: %.4f at epoch: %d. <<<" % \
(min_loss, min_loss_epoch)
#### Save elastic net model!! ####
if final_model != "NOSAVE":
with open(final_model, 'w') as final_state:
pickle.dump(sgd_clf, final_state)
def run(img_folder,
dl_state,
img_extension='dcm',
img_height=1024,
img_scale=255.,
equalize_hist=False,
featurewise_center=False,
featurewise_mean=91.6,
neg_vs_pos_ratio=1.,
net='vgg19',
batch_size=128,
patch_size=256,
stride=8,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
out='./modelState/prob_heatmap.pkl',
predicted_subj_file=None,
add_subjs=500):
'''Sweep mammograms with trained DL model to create prob heatmaps
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_labs = np.array(subj_labs)
print "Found %d subjests" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
if predicted_subj_file is not None:
predicted_subjs = np.load(predicted_subj_file)
subj_list = np.setdiff1d(subj_list, predicted_subjs)
subj_list = subj_list[:add_subjs]
print "Will predict additional %d subjects" % (len(subj_list))
elif neg_vs_pos_ratio is not None:
subj_list, subj_labs = DMMetaManager.subset_subj_list(
subj_list, subj_labs, neg_vs_pos_ratio, random_seed)
subj_labs = np.array(subj_labs)
print "After subsetting, there are %d subjects" % (len(subj_list))
print "cancer patients=%d, normal patients=%d" \
% ((subj_labs==1).sum(), (subj_labs==0).sum())
# Get exam lists.
# >>>> Debug <<<< #
# subj_list = subj_list[:2]
# >>>> Debug <<<< #
print "Get flattened exam list"
exam_list = meta_man.get_flatten_exam_list(subj_list, cc_mlo_only=True)
exam_labs = meta_man.exam_labs(exam_list)
exam_labs = np.array(exam_labs)
print "positive exams=%d, negative exams=%d" \
% ((exam_labs==1).sum(), (exam_labs==0).sum())
sys.stdout.flush()
# Load DL model.
print "Load patch classifier:", dl_state
sys.stdout.flush()
dl_model = load_model(dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
})
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model, _ = make_parallel(dl_model, gpu_count)
parallelized = True
else:
parallelized = False
# Load preprocess function.
if featurewise_center:
preprocess_input = None
else:
print "Load preprocess function for net:", net
if net == 'resnet50':
from keras.applications.resnet50 import preprocess_input
elif net == 'vgg16':
from keras.applications.vgg16 import preprocess_input
elif net == 'vgg19':
from keras.applications.vgg19 import preprocess_input
elif net == 'xception':
from keras.applications.xception import preprocess_input
elif net == 'inception':
from keras.applications.inception_v3 import preprocess_input
else:
raise Exception("Pretrained model is not available: " + net)
# Sweep the whole images and classify patches.
print "Generate prob heatmaps for exam list"
sys.stdout.flush()
heatmap_dat_list = []
for i, e in enumerate(exam_list):
dat = (e[0], e[1], {
'L': {
'cancer': e[2]['L']['cancer']
},
'R': {
'cancer': e[2]['R']['cancer']
}
})
dat[2]['L']['CC'] = get_prob_heatmap(
e[2]['L']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['L']['MLO'] = get_prob_heatmap(
e[2]['L']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['R']['CC'] = get_prob_heatmap(
e[2]['R']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
dat[2]['R']['MLO'] = get_prob_heatmap(
e[2]['R']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
heatmap_dat_list.append(dat)
print "processed %d/%d exams" % (i + 1, len(exam_list))
sys.stdout.flush()
### DEBUG ###
# if i >= 1:
# break
### DEBUG ###
print "Done."
# Save the result.
print "Saving result to external files.",
sys.stdout.flush()
pickle.dump(heatmap_dat_list, open(out, 'w'))
print "Done."
def run(img_folder,
dl_state,
img_extension='dcm',
img_height=1024,
img_scale=4095,
val_size=.2,
neg_vs_pos_ratio=10.,
do_featurewise_norm=True,
featurewise_mean=873.6,
featurewise_std=739.3,
img_per_batch=2,
roi_per_img=32,
roi_size=(256, 256),
low_int_threshold=.05,
blob_min_area=3,
blob_min_int=.5,
blob_max_int=.85,
blob_th_step=10,
layer_name=['flatten_1', 'dense_1'],
layer_index=None,
roi_state=None,
roi_clf_bs=32,
pc_components=.95,
pc_whiten=True,
nb_words=[512],
km_max_iter=100,
km_bs=1000,
km_patience=20,
km_init=10,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
pca_km_states='./modelState/dlrepr_pca_km_models.pkl',
bow_train_out='./modelState/bow_dat_train.pkl',
bow_test_out='./modelState/bow_dat_test.pkl'):
'''Calculate bag of deep visual words count matrix for all breasts
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = RandomState(random_seed) # an rng used across board.
# Load and split image and label lists.
meta_man = DMMetaManager(exam_tsv=exam_tsv,
img_tsv=img_tsv,
img_folder=img_folder,
img_extension=img_extension)
subj_list, subj_labs = meta_man.get_subj_labs()
subj_train, subj_test, labs_train, labs_test = train_test_split(
subj_list,
subj_labs,
test_size=val_size,
stratify=subj_labs,
random_state=random_seed)
if neg_vs_pos_ratio is not None:
def subset_subj(subj, labs):
subj = np.array(subj)
labs = np.array(labs)
pos_idx = np.where(labs == 1)[0]
neg_idx = np.where(labs == 0)[0]
nb_neg_desired = int(len(pos_idx) * neg_vs_pos_ratio)
if nb_neg_desired >= len(neg_idx):
return subj.tolist()
else:
neg_chosen = rng.choice(neg_idx, nb_neg_desired, replace=False)
subset_idx = np.concatenate([pos_idx, neg_chosen])
return subj[subset_idx].tolist()
subj_train = subset_subj(subj_train, labs_train)
subj_test = subset_subj(subj_test, labs_test)
img_list, lab_list = meta_man.get_flatten_img_list(subj_train)
lab_list = np.array(lab_list)
print "Train set - Nb of positive images: %d, Nb of negative images: %d" \
% ( (lab_list==1).sum(), (lab_list==0).sum())
sys.stdout.flush()
# Create image generator for ROIs for representation extraction.
print "Create an image generator for ROIs"
sys.stdout.flush()
if do_featurewise_norm:
imgen = DMImageDataGenerator(featurewise_center=True,
featurewise_std_normalization=True)
imgen.mean = featurewise_mean
imgen.std = featurewise_std
else:
imgen = DMImageDataGenerator(samplewise_center=True,
samplewise_std_normalization=True)
# Load ROI classifier.
if roi_state is not None:
print "Load ROI classifier"
sys.stdout.flush()
roi_clf = load_model(roi_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
})
graph = tf.get_default_graph()
else:
roi_clf = None
graph = None
# Create ROI generators for pos and neg images separately.
print "Create ROI generators for pos and neg images"
sys.stdout.flush()
roi_generator = imgen.flow_from_candid_roi(
img_list,
target_height=img_height,
target_scale=img_scale,
class_mode=None,
validation_mode=True,
img_per_batch=img_per_batch,
roi_per_img=roi_per_img,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
tf_graph=graph,
roi_clf=roi_clf,
clf_bs=roi_clf_bs,
return_sample_weight=False,
seed=random_seed)
# Generate image patches and extract their DL representations.
print "Load DL representation model"
sys.stdout.flush()
dlrepr_model = DLRepr(dl_state,
custom_objects={
'sensitivity': dmm.sensitivity,
'specificity': dmm.specificity
},
layer_name=layer_name,
layer_index=layer_index)
last_output_size = dlrepr_model.get_output_shape()[-1][-1]
if last_output_size != 3 and last_output_size != 1:
raise Exception("The last output must be prob outputs (size=3 or 1)")
nb_tot_samples = len(img_list) * roi_per_img
print "Extract ROIs from pos and neg images"
sys.stdout.flush()
pred = dlrepr_model.predict_generator(roi_generator,
val_samples=nb_tot_samples)
for i, d in enumerate(pred):
print "Shape of representation/output data %d:" % (i), d.shape
sys.stdout.flush()
# Flatten feature maps, e.g. an 8x8 feature map will become a 64-d vector.
pred = [d.reshape((-1, d.shape[-1])) for d in pred]
for i, d in enumerate(pred):
print "Shape of flattened data %d:" % (i), d.shape
sys.stdout.flush()
# Split representations and prob outputs.
dl_repr = pred[0]
prob_out = pred[1]
if prob_out.shape[1] == 3:
prob_out = prob_out[:, 1] # pos class.
prob_out = prob_out.reshape((len(img_list), -1))
print "Reshape prob output to:", prob_out.shape
sys.stdout.flush()
# Use PCA to reduce dimension of the representation data.
if pc_components is not None:
print "Start PCA dimension reduction on DL representation"
sys.stdout.flush()
pca = PCA(n_components=pc_components, whiten=pc_whiten)
pca.fit(dl_repr)
print "Nb of PCA components:", pca.n_components_
print "Total explained variance ratio: %.4f" % \
(pca.explained_variance_ratio_.sum())
dl_repr_pca = pca.transform(dl_repr)
print "Shape of transformed representation data:", dl_repr_pca.shape
sys.stdout.flush()
else:
pca = None
# Use K-means to create a codebook for deep visual words.
print "Start K-means training on DL representation"
sys.stdout.flush()
clf_list = []
clust_list = []
# Shuffling indices for mini-batches learning.
perm_idx = rng.permutation(len(dl_repr))
for n in nb_words:
print "Train K-means with %d cluster centers" % (n)
sys.stdout.flush()
clf = MiniBatchKMeans(n_clusters=n,
init='k-means++',
max_iter=km_max_iter,
batch_size=km_bs,
compute_labels=True,
random_state=random_seed,
tol=0.0,
max_no_improvement=km_patience,
init_size=None,
n_init=km_init,
reassignment_ratio=0.01,
verbose=0)
clf.fit(dl_repr[perm_idx])
clf_list.append(clf)
clust = np.zeros_like(clf.labels_)
clust[perm_idx] = clf.labels_
clust = clust.reshape((len(img_list), -1))
clust_list.append(clust)
if pca is not None:
print "Start K-means training on transformed representation"
sys.stdout.flush()
clf_list_pca = []
clust_list_pca = []
# Shuffling indices for mini-batches learning.
perm_idx = rng.permutation(len(dl_repr_pca))
for n in nb_words:
print "Train K-means with %d cluster centers" % (n)
sys.stdout.flush()
clf = MiniBatchKMeans(n_clusters=n,
init='k-means++',
max_iter=km_max_iter,
batch_size=km_bs,
compute_labels=True,
random_state=random_seed,
tol=0.0,
max_no_improvement=km_patience,
init_size=None,
n_init=km_init,
reassignment_ratio=0.01,
verbose=0)
clf.fit(dl_repr_pca[perm_idx])
clf_list_pca.append(clf)
clust = np.zeros_like(clf.labels_)
clust[perm_idx] = clf.labels_
clust = clust.reshape((len(img_list), -1))
clust_list_pca.append(clust)
# Read exam lists.
exam_train = meta_man.get_flatten_exam_list(subj_train,
flatten_img_list=True)
exam_test = meta_man.get_flatten_exam_list(subj_test,
flatten_img_list=True)
exam_labs_train = np.array(meta_man.exam_labs(exam_train))
exam_labs_test = np.array(meta_man.exam_labs(exam_test))
nb_pos_exams_train = (exam_labs_train == 1).sum()
nb_neg_exams_train = (exam_labs_train == 0).sum()
nb_pos_exams_test = (exam_labs_test == 1).sum()
nb_neg_exams_test = (exam_labs_test == 0).sum()
print "Train set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_train, nb_neg_exams_train)
print "Test set - Nb of pos exams: %d, Nb of neg exams: %d" % \
(nb_pos_exams_test, nb_neg_exams_test)
# Do BoW counts for each breast.
print "BoW counting for train exam list"
sys.stdout.flush()
bow_dat_train = get_exam_bow_dat(exam_train,
nb_words,
roi_per_img,
img_list=img_list,
prob_out=prob_out,
clust_list=clust_list)
for i, d in enumerate(bow_dat_train[1]):
print "Shape of train BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
print "BoW counting for test exam list"
sys.stdout.flush()
bow_dat_test = get_exam_bow_dat(exam_test,
nb_words,
roi_per_img,
imgen=imgen,
clf_list=clf_list,
transformer=None,
target_height=img_height,
target_scale=img_scale,
img_per_batch=img_per_batch,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
seed=random_seed,
dlrepr_model=dlrepr_model)
for i, d in enumerate(bow_dat_test[1]):
print "Shape of test BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
if pca is not None:
print "== Do same BoW counting on PCA transformed data =="
print "BoW counting for train exam list"
sys.stdout.flush()
bow_dat_train_pca = get_exam_bow_dat(exam_train,
nb_words,
roi_per_img,
img_list=img_list,
prob_out=prob_out,
clust_list=clust_list_pca)
for i, d in enumerate(bow_dat_train_pca[1]):
print "Shape of train BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
print "BoW counting for test exam list"
sys.stdout.flush()
bow_dat_test_pca = get_exam_bow_dat(
exam_test,
nb_words,
roi_per_img,
imgen=imgen,
clf_list=clf_list_pca,
transformer=pca,
target_height=img_height,
target_scale=img_scale,
img_per_batch=img_per_batch,
roi_size=roi_size,
low_int_threshold=low_int_threshold,
blob_min_area=blob_min_area,
blob_min_int=blob_min_int,
blob_max_int=blob_max_int,
blob_th_step=blob_th_step,
seed=random_seed,
dlrepr_model=dlrepr_model)
for i, d in enumerate(bow_dat_test_pca[1]):
print "Shape of test BoW matrix %d:" % (i), d.shape
sys.stdout.flush()
# Save K-means model and BoW count data.
if pca is None:
pickle.dump(clf_list, open(pca_km_states, 'w'))
pickle.dump(bow_dat_train, open(bow_train_out, 'w'))
pickle.dump(bow_dat_test, open(bow_test_out, 'w'))
else:
pickle.dump((pca, clf_list), open(pca_km_states, 'w'))
pickle.dump((bow_dat_train, bow_dat_train_pca),
open(bow_train_out, 'w'))
pickle.dump((bow_dat_test, bow_dat_test_pca), open(bow_test_out, 'w'))
print "Done."
def run(img_folder,
dl_state,
clf_info_state,
img_extension='dcm',
img_height=4096,
img_scale=255.,
equalize_hist=False,
featurewise_center=False,
featurewise_mean=91.6,
net='resnet50',
batch_size=64,
patch_size=256,
stride=64,
exam_tsv='./metadata/exams_metadata.tsv',
img_tsv='./metadata/images_crosswalk.tsv',
validation_mode=False,
use_mean=False,
out_pred='./output/predictions.tsv',
progress='./progress.txt'):
'''Run SC1 inference using prob heatmaps
'''
# Read some env variables.
random_seed = int(os.getenv('RANDOM_SEED', 12345))
rng = np.random.RandomState(random_seed) # an rng used across board.
gpu_count = int(os.getenv('NUM_GPU_DEVICES', 1))
# Setup data generator for inference.
meta_man = DMMetaManager(img_tsv=img_tsv,
exam_tsv=exam_tsv,
img_folder=img_folder,
img_extension=img_extension)
if validation_mode:
exam_list = meta_man.get_flatten_exam_list(cc_mlo_only=True)
exam_labs = meta_man.exam_labs(exam_list)
exam_labs = np.array(exam_labs)
print "positive exams=%d, negative exams=%d" \
% ((exam_labs==1).sum(), (exam_labs==0).sum())
sys.stdout.flush()
else:
exam_list = meta_man.get_last_exam_list(cc_mlo_only=True)
exam_labs = None
# Load DL model and classifiers.
print "Load patch classifier:", dl_state
sys.stdout.flush()
dl_model = load_model(dl_state)
if gpu_count > 1:
print "Make the model parallel on %d GPUs" % (gpu_count)
sys.stdout.flush()
dl_model, _ = make_parallel(dl_model, gpu_count)
parallelized = True
else:
parallelized = False
feature_name, nb_phm, cutoff_list, k, clf_list = \
pickle.load(open(clf_info_state))
# Load preprocess function.
if featurewise_center:
preprocess_input = None
else:
print "Load preprocess function for net:", net
if net == 'resnet50':
from keras.applications.resnet50 import preprocess_input
elif net == 'vgg16':
from keras.applications.vgg16 import preprocess_input
elif net == 'vgg19':
from keras.applications.vgg19 import preprocess_input
elif net == 'xception':
from keras.applications.xception import preprocess_input
elif net == 'inception':
from keras.applications.inception_v3 import preprocess_input
else:
raise Exception("Pretrained model is not available: " + net)
# Print header.
fout = open(out_pred, 'w')
if validation_mode:
fout.write(dminfer.INFER_HEADER_VAL)
else:
fout.write(dminfer.INFER_HEADER)
print "Start inference for exam list"
sys.stdout.flush()
for i, e in enumerate(exam_list):
### DEBUG ###
# if i >= 3:
# break
### DEBUG ###
subj = e[0]
exam_idx = e[1]
if validation_mode:
left_cancer = e[2]['L']['cancer']
right_cancer = e[2]['R']['cancer']
left_cancer = 0 if np.isnan(left_cancer) else left_cancer
right_cancer = 0 if np.isnan(right_cancer) else right_cancer
try:
left_cc_phms = get_prob_heatmap(
e[2]['L']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
except:
left_cc_phms = [None]
try:
left_mlo_phms = get_prob_heatmap(
e[2]['L']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
except:
left_mlo_phms = [None]
try:
right_cc_phms = get_prob_heatmap(
e[2]['R']['CC'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
except:
right_cc_phms = [None]
try:
right_mlo_phms = get_prob_heatmap(
e[2]['R']['MLO'],
img_height,
img_scale,
patch_size,
stride,
dl_model,
batch_size,
featurewise_center=featurewise_center,
featurewise_mean=featurewise_mean,
preprocess=preprocess_input,
parallelized=parallelized,
equalize_hist=equalize_hist)
except:
right_mlo_phms = [None]
try:
left_pred = dminfer.make_pred_case(left_cc_phms,
left_mlo_phms,
feature_name,
cutoff_list,
clf_list,
k=k,
nb_phm=nb_phm,
use_mean=use_mean)
except:
print "Exception in predicting left breast" + \
" for subj:", subj, "exam:", exam_idx
sys.stdout.flush()
left_pred = 0.
try:
right_pred = dminfer.make_pred_case(right_cc_phms,
right_mlo_phms,
feature_name,
cutoff_list,
clf_list,
k=k,
nb_phm=nb_phm,
use_mean=use_mean)
except:
print "Exception in predicting right breast" + \
" for subj:", subj, "exam:", exam_idx
sys.stdout.flush()
right_pred = 0.
if validation_mode:
fout.write("%s\t%s\tL\t%f\t%f\n" % \
(str(subj), str(exam_idx), left_pred, left_cancer))
fout.write("%s\t%s\tR\t%f\t%f\n" % \
(str(subj), str(exam_idx), right_pred, right_cancer))
fout.flush()
else:
fout.write("%s\tL\t%f\n" % (str(subj), left_pred))
fout.write("%s\tR\t%f\n" % (str(subj), right_pred))
fout.flush()
print "processed %d/%d exams" % (i + 1, len(exam_list))
sys.stdout.flush()
with open(progress, 'w') as fpro:
fpro.write("%f\n" % ((i + 1.) / len(exam_list)))
print "Done."
fout.close()