def run_liver(istrain=False,
model_name='unet',
modelcheckpoint='cache/liver/model/unet.hdf5',
model_pretrain='cache/liver/model/weight_unet_gen_tf.h5',
batch_size=1,
nb_epoch=500,
is_datagen=False,
channel=5):
reader = DataSetVolumn()
seg = SegmentationBatch(model_name, modelcheckpoint)
if istrain:
current_dir = os.path.dirname(os.path.realpath(__file__))
weights_path = os.path.expanduser(
os.path.join(current_dir,
model_pretrain)) if model_pretrain else None
seg.train(reader,
weights_path=weights_path,
batch_size=batch_size,
nb_epoch=nb_epoch,
is_datagen=is_datagen,
channel=channel)
testdata = reader.load_testdata_channel(channel)
metrics_testdata = []
for imagefile, data in testdata.iteritems():
X_test, y_test = data
predicts = seg.predict(X_test, batch_size=batch_size)
pprint(Metrics.all(y_test, predicts))
metrics_testdata.append((imagefile, Metrics.all(y_test, predicts)))
result = {
'acc':
sum([metrics['acc'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'dice':
sum([metrics['dice'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'jacc':
sum([metrics['jacc'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'sensitivity':
sum([
metrics['sensitivity']
for imagefile, metrics in metrics_testdata
]) / len(metrics_testdata),
'specificity':
sum([
metrics['specificity']
for imagefile, metrics in metrics_testdata
]) / len(metrics_testdata)
}
print('the average metrics case by case')
pprint(result)
return (X_test, y_test, predicts)
def run_skin(istrain,
model_name,
modelcheckpoint,
model_pretrain=None,
batch_size=32,
nb_epoch=200,
is_datagen=False):
print('-' * 30)
print('Loading data...')
print('-' * 30)
X_train, X_test, y_train, y_test = SkinData.load_from_npy(
images_npy='cache/skin/datasets/images_224_224_tf.npy',
masks_npy='cache/skin/datasets/masks_224_224_tf.npy')
seg = Segmentation(model_name, modelcheckpoint)
if istrain:
seg.train(X_train,
y_train, (X_test, y_test),
weights_path=Segmentation.absolute_path(model_pretrain),
batch_size=batch_size,
nb_epoch=nb_epoch,
is_datagen=is_datagen)
predicts = seg.predict(X_test, batch_size=batch_size)
pprint(Metrics.all(y_test, predicts))
return (X_test, y_test, predicts)
def run_breast(
istrain,
model_name,
modelcheckpoint,
model_pretrain=None,
batch_size=32,
nb_epoch=500,
is_datagen=False,
images_npy='cache/breast/datasets/images_crop_224_224_tf.npy',
masks_npy='cache/breast/datasets/masks_crop_224_224_tf.npy'):
print('-' * 30)
print('Loading data...')
print('-' * 30)
X_train, X_test, y_train, y_test = BreastData.load_from_npy(
images_npy, masks_npy, test_size=.1)
print(
'Done Loading train with shape {0} and test with shape {1}'.format(
X_train.shape, X_test.shape))
# shape = y_train.shape
# y_flat = np.reshape(y_train, shape[0]*shape[1]*shape[2]*shape[3]).astype('uint8')
# maxvalue, minvalue = np.max(y_flat), np.min(y_flat)
# cw = class_weight.compute_class_weight('balanced', np.unique(y_flat), y_flat)
seg = Segmentation(model_name, modelcheckpoint)
if istrain:
seg.train(X_train,
y_train, (X_test, y_test),
weights_path=Segmentation.absolute_path(model_pretrain),
batch_size=batch_size,
nb_epoch=nb_epoch,
is_datagen=is_datagen)
predicts = seg.predict(X_test, batch_size=batch_size)
pprint(Metrics.all(y_test, predicts))
return (X_test, y_test, predicts)
def run_liver(istrain,
model_name,
modelcheckpoint,
model_pretrain=None,
batch_size=1,
nb_epoch=500,
is_datagen=False,
isliver=True,
image_size=(512, 512)):
reader = DataSetVolumn(
image_size=image_size) if isliver else TumorVolumn(
image_size=image_size)
seg = Segmentation(model_name, modelcheckpoint)
if istrain:
X_train, y_train = reader.load_traindata()
seg.train(X_train,
y_train,
reader.validation_data(),
weights_path=Segmentation.absolute_path(model_pretrain),
batch_size=batch_size,
nb_epoch=nb_epoch,
is_datagen=is_datagen)
testdata = reader.load_testdata()
metrics_testdata = []
for imagefile, data in testdata.iteritems():
X_test, y_test = data
predicts = seg.predict(X_test, batch_size=batch_size)
def save(liver):
volumefile = 'volumn-' + imagefile.split('-')[1] + '.npy'
maskfile = 'segmentation-' + imagefile.split('-')[1] + '.npy'
predictfile = 'predict-' + imagefile.split('-')[1] + '.npy'
np.save(
os.path.join('cache/{0}/result'.format(liver), volumefile),
X_test)
np.save(
os.path.join('cache/{0}/result'.format(liver), maskfile),
y_test)
np.save(
os.path.join('cache/{0}/result'.format(liver),
predictfile), predicts)
# save('liver') if isliver else save('tumor')
def tumor2noscale(imagefile, predicts):
if not isliver:
boxs, images, masks = reader.get_data(imagefile)
masks[masks < 2] = 0
masks /= 2
imagesize = (512, 512)
predicts_new = np.zeros(
(predicts.shape[0], imagesize[0], imagesize[1], 1),
dtype=predicts.dtype)
for i, predict in enumerate(predicts):
predicts_new[i, :, :,
0] = cv.resize(predict[:, :, 0],
imagesize)
predicts_new = to_category(predicts_new)
predicts_new = back2noscale(boxs, predicts_new)
# viewSequence(masks)
# viewSequence(predicts_new)
y_test, predicts = masks, predicts_new
tumor2noscale(imagefile, predicts)
pprint(Metrics.all(y_test, predicts))
metrics_testdata.append((imagefile, Metrics.all(y_test, predicts)))
result = {
'acc':
sum([metrics['acc'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'dice':
sum([metrics['dice'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'jacc':
sum([metrics['jacc'] for imagefile, metrics in metrics_testdata]) /
len(metrics_testdata),
'sensitivity':
sum([
metrics['sensitivity']
for imagefile, metrics in metrics_testdata
]) / len(metrics_testdata),
'specificity':
sum([
metrics['specificity']
for imagefile, metrics in metrics_testdata
]) / len(metrics_testdata)
}
print('the average metrics case by case')
pprint(result)
return (X_test, y_test, predicts)