def exp_eval_ots_lidc_jsrt(args):
# load LIDC & JSRT-positives data
imgs_tr, blobs_tr = lidc.load()
pred_blobs_tr = detect.read_blobs('data/{}-lidc-pred-blobs.pkl'.format(
args.detector))
masks_tr = np.load('data/aam-lidc-pred-masks.npy')
imgs_te, blobs_te = jsrt.load(set_name='jsrt140p')
pred_blobs_te = detect.read_blobs('data/{}-jsrt140p-pred-blobs.pkl'.format(
args.detector))
masks_te = np.load('data/aam-jsrt140p-pred-masks.npy')
rois_tr = lnd.create_rois(imgs_tr, masks_tr, pred_blobs_tr, args)
rois_te = lnd.create_rois(imgs_te, masks_te, pred_blobs_te, args)
# Create rois dataset
rois_tr, Y_tr, _, _ = neural.create_train_test_sets(
blobs_tr, pred_blobs_tr, rois_tr, None, None, None)
generator = augment.get_default_generator((args.roi_size, args.roi_size))
rois_tr, Y_tr = augment.balance_and_perturb(rois_tr, Y_tr, generator)
range_tr = np.max(rois_tr), np.min(rois_tr)
print "range {}".format(range_tr)
# Extract features
network = VGG16(mode='ots-feat', pool_layer=args.pool_layer)
feats_tr = extract_convfeats(network, rois_tr, range_tr)
feats_te = extract_convfeats(network, rois_te, range_tr)
np.save('data/{}-lidc-feats.npy'.format(args.detector, fold_idx), feats_tr)
np.save('data/{}-jsrt140p-feats.npy'.format(args.detector, fold_idx),
feats_te)
# Eval classifier
clf = LinearSVC(C=args.svm_c)
froc = evaluate_classifier(clf, feats_tr, Y_tr, blobs_te, pred_blobs_te,
feats_te)
def visual_results_jsrt_only(model_name, args):
print "Visual results for model {} JSRT only".format(model_name)
imgs, blobs = jsrt.load(set_name='jsrt140p')
pred_blobs = detect.read_blobs('data/{}-jsrt140p-pred-blobs.pkl'.format(
args.detector))
masks = np.load('data/aam-jsrt140p-pred-masks.npy')
rois = create_rois(imgs, masks, pred_blobs, args)
folds = KFold(n_splits=5, shuffle=True,
random_state=util.FOLDS_SEED).split(imgs)
fold_idx = 0
for tr, te in folds:
model.load('data/' + model.name + '.fold-{}'.format(fold_idx + 1))
model = neural.create_network(model_name, args,
(1, args.roi_size, args.roi_size))
X_tr, Y_tr, X_te, Y_te = neural.create_train_test_sets(
real_blobs_tr, pred_blobs_tr, rois_tr, real_blobs_te,
pred_blobs_te, rois_te)
print 'load weights {}'.format(model.name)
model.network.load_weights('data/{}_weights.h5'.format(model.name))
# FIX: remove and add zmuv mean and zmuv std no Preprocessor augment.py
if not hasattr(model.preprocessor, 'zmuv_mean'):
model.preprocessor.fit(X_tr, Y_tr)
model.save('data/' + model.name)
pred_blobs_te, probs_te = neural.predict_proba(model, pred_blobs_te,
rois_te)
util.save_rois_with_probs(rois_te, probs_te)
fold_idx += 1
def model_output(model_name, args):
print "Model Outputs"
imgs, blobs = jsrt.load(set_name='jsrt140p')
pred_blobs = detect.read_blobs('data/{}-jsrt140p-pred-blobs.pkl'.format(
args.detector))
masks = np.load('data/aam-jsrt140p-pred-masks.npy')
rois = create_rois(imgs, masks, pred_blobs, args)
folds = KFold(n_splits=5, shuffle=True,
random_state=util.FOLDS_SEED).split(imgs)
fold_idx = 0
frocs = []
legends = ['Fold {}'.format(i + 1) for i in range(5)]
index = np.array(range(len(imgs)))
for tr, te in folds:
X_tr, Y_tr, _, _ = neural.create_train_test_sets(
blobs[tr], pred_blobs[tr], rois[tr], blobs[te], pred_blobs[te],
rois[te])
model = neural.create_network(model_name, args,
(1, args.roi_size, args.roi_size))
model.name = model.name + '-{}-lidc.fold-{}'.format(
args.detector, fold_idx + 1)
model.network.load_weights('data/{}_weights.h5'.format(model.name))
if not hasattr(model.preprocessor, 'zmuv_mean'):
model.preprocessor.fit(X_tr, Y_tr)
print "Predict ..."
pred_blobs_te, probs_te, rois_te = neural.predict_proba(
model, pred_blobs[te], rois[te])
print "Save ..."
eval.save_outputs(imgs[te], blobs[te], pred_blobs_te, probs_te,
rois_te, index[te])
def evaluate_model(model,
real_blobs_tr,
pred_blobs_tr,
rois_tr,
real_blobs_te,
pred_blobs_te,
rois_te,
load_model=False):
X_tr, Y_tr, X_te, Y_te = neural.create_train_test_sets(
real_blobs_tr, pred_blobs_tr, rois_tr, real_blobs_te, pred_blobs_te,
rois_te)
if load_model == True:
print 'load weights {}'.format(model.name)
model.network.load_weights('data/{}_weights.h5'.format(model.name))
# FIX: remove and add zmuv mean and zmuv std no Preprocessor augment.py
if not hasattr(model.preprocessor, 'zmuv_mean'):
model.preprocessor.fit(X_tr, Y_tr)
else:
_ = model.fit(X_tr, Y_tr, X_te, Y_te)
model.save('data/' + model.name)
pred_blobs_te, probs_te, _ = neural.predict_proba(model, pred_blobs_te,
rois_te)
return eval.froc(real_blobs_te, pred_blobs_te, probs_te)
def save_rois(args):
imgs_tr, blobs_tr = lidc.load(pts=False)
pred_blobs_tr = detect.read_blobs('data/sbf-aam-lidc-pred-blobs.pkl')
masks_tr = np.load('data/aam-lidc-pred-masks.npy')
imgs_te, blobs_te = jsrt.load(set_name='jsrt140p')
pred_blobs_te = detect.read_blobs('data/sbf-aam-jsrt140p-pred-blobs.pkl')
masks_te = np.load('data/aam-jsrt140p-pred-masks.npy')
rois_tr = create_rois(imgs_tr,
masks_tr,
pred_blobs_tr,
args,
real_blobs=blobs_tr)
rois_te = create_rois(imgs_te,
masks_te,
pred_blobs_te,
args,
real_blobs=blobs_te)
X_tr, Y_tr, X_te, Y_te = neural.create_train_test_sets(
blobs_tr, pred_blobs_tr, rois_tr, blobs_te, pred_blobs_te, rois_te)
X_tr, Y_tr = util.split_data_pos_neg(X_tr, Y_tr)
X_te, Y_te = util.split_data_pos_neg(X_te, Y_te)
X_pos = X_tr[0]
idx = np.random.randint(0, len(X_tr[1]), len(X_pos))
X_neg = X_tr[1][idx]
print len(X_pos), len(X_neg)
for i in range(len(X_pos)):
util.imwrite('data/lidc/roi{}p.jpg'.format(i), X_pos[i][0])
np.save('data/lidc/roi{}p.npy'.format(i), X_pos[i])
util.imwrite('data/lidc/roi{}n.jpg'.format(i), X_neg[i][0])
np.save('data/lidc/roi{}n.npy'.format(i), X_neg[i])
X_pos = X_te[0]
idx = np.random.randint(0, len(X_te[1]), len(X_pos))
X_neg = X_te[1][idx]
print len(X_pos), len(X_neg)
for i in range(len(X_pos)):
util.imwrite('data/jsrt140/roi{}p.jpg'.format(i), X_pos[i][0])
np.save('data/jsrt140/roi{}p.npy'.format(i), X_pos[i])
util.imwrite('data/jsrt140/roi{}n.jpg'.format(i), X_neg[i][0])
np.save('data/jsrt140/roi{}n.npy'.format(i), X_neg[i])
def train_with_feature_set_keras(self, feats_tr, pred_blobs_tr, real_blobs_tr,
feats_test=None, pred_blobs_test=None, real_blobs_test=None,
model='shallow_1', model_suffix=None, network_init=None):
print("{} {} {} {} {} {}".format(len(feats_tr), len(pred_blobs_tr), len(real_blobs_tr), len(feats_test), len(pred_blobs_test), len(real_blobs_test)))
X_train, Y_train, X_test, Y_test = neural.create_train_test_sets(feats_tr, pred_blobs_tr, real_blobs_tr,
feats_test, pred_blobs_test, real_blobs_test, streams=self.streams )
print "X_train shape {}".format(X_train.shape)
self.network = neural.create_network(model, (X_train.shape[1], self.roi_size, self.roi_size), self.streams)
if network_init is not None:
if self.args.transfer:
self.load_cnn_weights('data/{}_{}'.format(network_init, model_suffix))
else:
self.load_cnn_weights(network_init)
name = 'data/{}_{}'.format(model, model_suffix)
history = self.network.fit(X_train, Y_train, X_test, Y_test, streams=(self.streams != 'none'), cropped_shape=(self.roi_size, self.roi_size), checkpoint_prefix=name, checkpoint_interval=2)
return history
def exp_eval_ots_jsrt_only(args):
# load LIDC & JSRT-positives data
network = VGG16(mode='ots-feat', pool_layer=args.pool_layer)
print "Model Evaluation Protocol 2"
imgs, blobs = jsrt.load(set_name='jsrt140p')
pred_blobs = detect.read_blobs('data/{}-jsrt140p-pred-blobs.pkl'.format(
args.detector))
masks = np.load('data/aam-jsrt140p-pred-masks.npy')
rois = lnd.create_rois(imgs, masks, pred_blobs, args)
folds = KFold(n_splits=5, shuffle=True,
random_state=util.FOLDS_SEED).split(imgs)
fold_idx = 0
frocs = []
legends = ['Fold {}'.format(i + 1) for i in range(5)]
for tr, te in folds:
# Eval classifier
rois_tr, Y_tr, _, _ = neural.create_train_test_sets(
blobs_tr, pred_blobs_tr, rois[tr], None, None, None)
generator = augment.get_default_generator(
(args.roi_size, args.roi_size))
rois_tr, Y_tr = augment.balance_and_perturb(rois_tr, Y_tr, generator)
clf = LinearSVC(C=args.svm_c)
froc = evaluate_classifier(clf, rois_tr, Y_tr, blobs[te],
pred_blobs[te], feats[te])
frocs.append(froc)
current_frocs = [eval.average_froc([froc_i]) for froc_i in frocs]
util.save_froc(current_frocs,
'data/lsvm-{}-jsrtonly-folds'.format(args.detector),
legends[:len(frocs)],
with_std=False)
fold_idx += 1
froc = eval.average_froc(frocs)
legends = ['Test FROC (JSRT positives)']
util.save_froc([froc],
'data/lsvm-{}-jsrtonly'.format(args.detector),
legends,
with_std=True)
def extract_features_from_convnet(args): # Load img, blobs and masks
imgs, blobs, paths = lidc.load(pts=True, set_name=args.ds_tr)
pred_blobs = detect.read_blobs('data/{}-lidc-pred-blobs.pkl'.format(
args.detector))
masks = np.load('data/aam-lidc-pred-masks.npy')
assert len(imgs) == len(masks) and len(pred_blobs) == len(masks)
# Load folds
folds = util.model_selection_folds(imgs)
# Create rois
rois = lnd.create_rois(imgs, masks, pred_blobs, args, real_blobs=blobs)
# Load model
network = VGG16(mode='ots-feat', pool_layer=args.pool_layer)
network.summary()
# Set up CV
frocs = []
legends = ['Fold {}'.format(i + 1) for i in range(util.NUM_VAL_FOLDS)]
fold_idx = 0
for tr, te in folds:
# TODO: apply extract convfeats funcs for tr and te sets
print "Fold {}".format(fold_idx + 1)
X_tr, Y_tr, _, _ = neural.create_train_test_sets(
blobs[tr], pred_blobs[tr], rois[tr], blobs[te], pred_blobs[te],
rois[te])
gc.collect()
generator = augment.get_default_generator(
(args.roi_size, args.roi_size))
X_tr, Y_tr = augment.balance_and_perturb(X_tr, Y_tr, generator)
gc.collect()
''''
counta = 0
countb = 0
count = 0
while counta < 10 and countb < 10:
if Y_tr[count][1] > 0 and counta < 10:
util.imshow("positives", X_tr[count][0], display_shape=(256, 256))
counta += 1
elif Y_tr[count][1] == 0.0 and countb < 10:
util.imshow("negatives", X_tr[count][0], display_shape=(256, 256))
countb += 1
count += 1
'''
range_tr = (X_tr.min(), X_tr.max())
print "Range {}".format(range_tr)
print "Extract feats on balanced tr set"
feats_tr = extract_convfeats(network, X_tr, range_tr)
save_features(
"data/vgg16-{}-{}-f{}-lidc-feats".format(args.pool_layer,
args.detector, fold_idx),
feats_tr, Y_tr)
gc.collect()
print "Extract feats on te set"
feats_te = extract_convfeats_from_rois(network, rois[te], range_tr)
print "Test feats to save shape {}".format(feats_te.shape)
np.save(
"data/vgg16-{}-{}-f{}-te-lidc-feats.npy".format(
args.pool_layer, args.detector, fold_idx), feats_te)
gc.collect()
fold_idx += 1
def froc_by_epochs(data,
blobs,
augmented_blobs,
rois,
folds,
network_model,
nb_epochs=30,
epoch_interval=2):
network_init = None
roi_size = 32
streams = 'none'
imgs = []
masks = []
for i in range(len(data)):
img, lung_mask = data.get(i, downsample=True)
sampled, lce, norm = preprocess.preprocess_hardie(img,
lung_mask,
downsample=True)
imgs.append([lce])
masks.append(lung_mask)
imgs = np.array(imgs)
masks = np.array(masks)
# Hardcoding blob set shapes
blobs2 = blobs
blobs = blobs.reshape((len(blobs), 3))
nb_checkpoints = int(nb_epochs / epoch_interval)
epochs = np.linspace(epoch_interval, nb_checkpoints * epoch_interval,
nb_checkpoints).astype(np.int)
av_frocs = []
names = []
aucs1 = []
aucs2 = []
for epoch in epochs:
frocs = []
fold = 1
for tr_idx, te_idx in folds:
print "Fold {} ...".format(fold)
X_train, Y_train, X_test, Y_test = neural.create_train_test_sets(
rois[tr_idx],
augmented_blobs[tr_idx],
blobs[tr_idx],
rois[te_idx],
augmented_blobs[te_idx],
blobs[te_idx],
streams=streams,
detector=True)
# load network
network = neural.create_network(network_model,
X_train.shape,
fold,
streams,
detector=False)
name = 'data/{}_fold_{}.epoch_{}'.format(network_model, fold,
epoch)
network.network.load_weights('{}_weights.h5'.format(name))
# open network on detector mode
detector_network = neural.create_network(network_model,
X_train.shape,
fold,
streams,
detector=True)
copy_weights(network, detector_network)
# evaluate network on test
blobs_te_pred, probs_te_pred = detect_with_network(
detector_network, imgs[te_idx], masks[te_idx], fold=fold)
froc = eval.froc(blobs2[te_idx], blobs_te_pred, probs_te_pred)
frocs.append(froc)
fold += 1
names.append('{}, epoch {}'.format(network_model, epoch))
ops = eval.average_froc(frocs, fppi_range)
av_frocs.append(ops)
aucs1.append(util.auc(ops, range(0, 60)))
aucs2.append(util.auc(ops, range(0, 40)))
util.save_auc(
np.array(range(1,
len(aucs1) + 1)) * epoch_interval, aucs1,
'data/{}-auc-0-60'.format(network_model))
util.save_auc(
np.array(range(1,
len(aucs2) + 1)) * epoch_interval, aucs2,
'data/{}-auc-0-40'.format(network_model))
return av_frocs, names
def eval_cnn_detector(data, blobs, augmented_blobs, rois, folds, model):
fold = 1
network_init = None
roi_size = 32
streams = 'none'
imgs = []
masks = []
for i in range(len(data)):
img, lung_mask = data.get(i, downsample=True)
sampled, lce, norm = preprocess.preprocess_hardie(img,
lung_mask,
downsample=True)
imgs.append([lce])
masks.append(lung_mask)
imgs = np.array(imgs)
masks = np.array(masks)
# Hardcoding blob set shapes
blobs2 = blobs
blobs = blobs.reshape((len(blobs), 3))
frocs = []
for tr_idx, te_idx in folds:
print "Fold {} ...".format(fold)
X_train, Y_train, X_test, Y_test = neural.create_train_test_sets(
rois[tr_idx],
augmented_blobs[tr_idx],
blobs[tr_idx],
rois[te_idx],
augmented_blobs[te_idx],
blobs[te_idx],
streams=streams,
detector=True)
network = neural.create_network(model,
X_train.shape,
fold,
streams,
detector=False)
if network_init is not None:
network.network.load_weights('data/{}_fold_{}_weights.h5'.format(
network_init, fold))
# save network
name = 'data/{}_fold_{}'.format(model, fold)
history = network.fit(X_train,
Y_train,
X_test,
Y_test,
streams=(streams != 'none'),
cropped_shape=(roi_size, roi_size),
checkpoint_prefix=name,
checkpoint_interval=2,
loss='mse')
network.save(name)
# open network on detector mode
network.network.summary()
detector_network = neural.create_network(model,
X_train.shape,
fold,
streams,
detector=True)
detector_network.network.summary()
copy_weights(network, detector_network)
#network.network.load_weights('{}_weights.h5'.format(name))
#network.load(name)
# evaluate network on test
blobs_te_pred, probs_te_pred = detect_with_network(detector_network,
imgs[te_idx],
masks[te_idx],
fold=fold)
froc = eval.froc(blobs2[te_idx], blobs_te_pred, probs_te_pred)
frocs.append(froc)
fold += 1
av_froc = eval.average_froc(frocs, fppi_range)
return av_froc
