def train_network(basis,
model_name,
epochs,
label_norm,
batch_size,
n_train_spectra,
metabolites,
export_acquisitions,
export_datatype,
n_validation_spectra=None):
if not n_validation_spectra:
n_validation_spectra = np.ceil(n_train_spectra * 0.2)
save_root = SAVE_ROOT
n_train_lw_samples = int(np.round(float(n_train_spectra) / len(basis)))
n_validation_lw_samples = int(
np.round(float(n_validation_spectra) / len(basis)))
train_datasets = []
for basi in basis:
train_dataset = Dataset()
train_dataset._name = 'train'
train_dataset.basis = basi
train_dataset.export_datatype = export_datatype
train_dataset.export_acquisitions = export_acquisitions
train_dataset.conc_normalisation = label_norm
train_dataset.export_nu = False
train_dataset.export_dss_peak = False
train_dataset.add_adc_noise = True
train_dataset.add_nu_noise = False
train_dataset.generate_dataset(metabolites, n_train_lw_samples)
train_datasets.append(train_dataset)
benchmark_basis, benchmark_basis_names = load_benchmark_datasets()
benchmark_datasets = []
for basi, name in zip(benchmark_basis, benchmark_basis_names):
td = Dataset()
td._name = 'benchmark_' + name
td.basis = basi
td.copy_settings(train_dataset, test_dataset=True)
td.copy_basis()
benchmark_datasets.append([td])
validation_datasets = []
for basi in basis:
validation_dataset = Dataset()
validation_dataset.basis = basi
validation_dataset.copy_settings(train_dataset)
validation_dataset._name = 'validation_multi_lw'
validation_dataset.generate_dataset(metabolites,
n_validation_lw_samples)
validation_datasets.append(validation_dataset)
multi_dataset_train_loop(
model_name,
epochs,
batch_size,
label_norm,
save_dir=save_root + 'models/',
train_datasets=train_datasets,
test_datasets=benchmark_datasets + [validation_datasets],
dataset_normalisation=train_dataset.conc_normalisation)
def quantify(ima_dir, model_dir, metabolites=None):
basis = Basis.load_dicom(ima_dir)
model = load_network(model_dir)
if metabolites:
# check to see if the network can actually quantify the requested metabolites
if not all([
m_name.lower() in [x.lower() for x in model.output_labels]
for m_name in metabolites
]):
raise Exception(
'Network is unable to quantify one or more metabolites suplied.'
'Network is able to quantify: ' + str(model.output_labels) +
'\n'
'Requested metabolites: ' + str(metabolites))
else:
metabolites = model.output_labels
# generate a dataset from the loaded dicoms
dataset = Dataset()
dataset._name = 'quantify'
dataset.basis = basis
dataset.copy_basis()
dataset.export_datatype = model.export_datatype
dataset.export_acquisitions = model.export_acquisitions
dataset.conc_normalisation = model.output_normalisation
dataset.export_nu = False
dataset.export_dss_peak = False
dataset.add_adc_noise = False
dataset.add_nu_noise = False
# export the dataset into the format we're looking
t_data, t_labels, mo_labels = dataset.export_to_keras(
model_labels=metabolites)
t_data, input_shape = reshape_data(t_data)
# run the model in test mode to display the loss and accuracy
model.evaluate(x=t_data, y=t_labels)
# quantify!
predictions = model.predict(t_data)
# trim the output data to match the metabolites of interest
cols_index = [[x.lower()
for x in model.output_labels].index(m_name.lower())
for m_name in metabolites]
predictions = predictions[:, cols_index]
# renormalise the output labels
predictions = normalise_labels(predictions, model.output_normalisation)
# get the spectra in the order they were exported in dataset.export_to_keras
# so we can match the predictions to the spectra
spectra = np.array(dataset.group_spectra_by_id())
# print the results table!
print('\n\nQuantifying %d MEGA-PRESS Spectra' % (len(spectra)))
print('This network can only quantify: ' + str(model.output_labels))
print('\tNetwork path: ' + model_dir)
print('\tDICOM path: ' + ima_dir + '\n\n')
for spec, prediction in zip(spectra, predictions):
if sum(spec[0].concentrations):
print('Spectra ID: ' + spec[0].id)
print('\t%-*s %s %s' % (20, 'Metabolite', 'Predicded', 'Actual'))
actual_concentrations = normalise_labels(
spec[0].concentrations, model.output_normalisation)
for p, a, m_name in zip(
prediction, actual_concentrations,
convert_mol_names(metabolites, mode='lengthen')):
print('\t%-*s %.6f %.6f' % (20, m_name, p, a))
print('\n')
else:
print('Spectra ID: ' + spec[0].id)
print('\t%-*s %s' % (20, 'Metabolite', 'Predicted'))
for p, m_name in zip(
prediction, convert_mol_names(metabolites,
mode='lengthen')):
print('\t%-*s %.6f' % (20, m_name, p))
print('\n')
