Memory efficient training for 3D brain image registration networks using PatchMorph.

Code ↔ Paper

2 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 2 matches

[1] § Method › PatchMorph › Baseline CNN architecture ↔ NICE-Net/ext/neuron/models.py, lines 805 to 952 · score 0.61 · batch norm, ReLu, CNN, convolution, dense, stride · Report a mismatch
[2] § Experiments › PatchMorph architecture › PatchMorph training ↔ libs/codes.py, lines 898 to 1025 · score 0.54 · AdamW, NCC loss, PyTorch, optimizer, batch, training · Report a mismatch

Paper

Loaded from Europe PMC by your browser; OSCR does not store it: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The paper is loaded by your browser, with JavaScript: read it at doi.org or on Europe PMC.

The authors' code

Python · 1,009 lines · 36 KB · MIT · 1 match

"""
tensorflow/keras utilities for the neuron project
If you use this code, please cite
Dalca AV, Guttag J, Sabuncu MR
Anatomical Priors in Convolutional Networks for Unsupervised Biomedical Segmentation,
CVPR 2018
Contact: adalca [at] csail [dot] mit [dot] edu
License: GPLv3
"""
import sys
from . import layers
# third party
import numpy as np
import tensorflow as tf
import keras
import keras.layers as KL
from keras.models import Model
import keras.backend as K
from keras.constraints import maxnorm
def dilation_net(nb_features,
input_shape, # input layer shape, vector of size ndims + 1(nb_channels)
nb_levels,
conv_size,
nb_labels,
name='dilation_net',
prefix=None,
feat_mult=1,
pool_size=2,
use_logp=True,
padding='same',
dilation_rate_mult=1,
activation='elu',
use_residuals=False,
final_pred_activation='softmax',
nb_conv_per_level=1,
add_prior_layer=False,
add_prior_layer_reg=0,
layer_nb_feats=None,
batch_norm=None):
return unet(nb_features,
input_shape, # input layer shape, vector of size ndims + 1(nb_channels)
nb_levels,
conv_size,
nb_labels,
name='unet',
prefix=None,
feat_mult=1,
pool_size=2,
use_logp=True,
padding='same',
activation='elu',
use_residuals=False,
dilation_rate_mult=dilation_rate_mult,
final_pred_activation='softmax',
nb_conv_per_level=1,
add_prior_layer=False,
add_prior_layer_reg=0,
layer_nb_feats=None,
batch_norm=None)
def unet(nb_features,
input_shape, # input layer shape, vector of size ndims + 1(nb_channels)
nb_levels,
conv_size,
nb_labels,
name='unet',
prefix=None,
feat_mult=1,
pool_size=2,
use_logp=True,
padding='same',
dilation_rate_mult=1,
activation='elu',
use_residuals=False,
final_pred_activation='softmax',
nb_conv_per_level=1,
add_prior_layer=False,
add_prior_layer_reg=0,
layer_nb_feats=None,
conv_dropout=0,
batch_norm=None):
"""
unet-style model with an overdose of parametrization
for U-net like architecture, we need to use Deconvolution3D.
However, this is not yet available (maybe soon, it's on a dev branch in github I believe)
Until then, we'll upsample and convolve.
TODO: Need to check that UpSampling3D actually does NN-upsampling!
"""
# naming
model_name = name
if prefix is None:
prefix = model_name
# volume size data
ndims = len(input_shape) - 1
if isinstance(pool_size, int):
pool_size = (pool_size,) * ndims
# get encoding model
enc_model = conv_enc(nb_features,
input_shape,
nb_levels,
conv_size,
name=model_name,
prefix=prefix,
feat_mult=feat_mult,
pool_size=pool_size,
padding=padding,
dilation_rate_mult=dilation_rate_mult,
activation=activation,
use_residuals=use_residuals,
nb_conv_per_level=nb_conv_per_level,
layer_nb_feats=layer_nb_feats,
conv_dropout=conv_dropout,
batch_norm=batch_norm)
# get decoder
# use_skip_connections=1 makes it a u-net
lnf = layer_nb_feats[(nb_levels * nb_conv_per_level):] if layer_nb_feats is not None else None
dec_model = conv_dec(nb_features,
None,
nb_levels,
conv_size,
nb_labels,
name=model_name,
prefix=prefix,
feat_mult=feat_mult,
pool_size=pool_size,
use_skip_connections=1,
padding=padding,
dilation_rate_mult=dilation_rate_mult,
activation=activation,
use_residuals=use_residuals,
final_pred_activation='linear' if add_prior_layer else final_pred_activation,
nb_conv_per_level=nb_conv_per_level,
batch_norm=batch_norm,
layer_nb_feats=lnf,
conv_dropout=conv_dropout,
input_model=enc_model)
final_model = dec_model
if add_prior_layer:
final_model = add_prior(dec_model,
[*input_shape[:-1], nb_labels],
name=model_name + '_prior',
use_logp=use_logp,
final_pred_activation=final_pred_activation,
add_prior_layer_reg=add_prior_layer_reg)
return final_model
def ae(nb_features,
input_shape,
nb_levels,
conv_size,
nb_labels,
enc_size,
name='ae',
prefix=None,
feat_mult=1,
pool_size=2,
padding='same',
activation='elu',
use_residuals=False,
nb_conv_per_level=1,
batch_norm=None,
enc_batch_norm=None,
ae_type='conv', # 'dense', or 'conv'
enc_lambda_layers=None,
add_prior_layer=False,
add_prior_layer_reg=0,
use_logp=True,
conv_dropout=0,
include_mu_shift_layer=False,
single_model=False, # whether to return a single model, or a tuple of models that can be stacked.
final_pred_activation='softmax',
do_vae=False):
"""
Convolutional Auto-Encoder.
Optionally Variational.
Optionally Dense middle layer
"Mostly" in that the inner encoding can be (optionally) constructed via dense features.
Parameters:
do_vae (bool): whether to do a variational auto-encoder or not.
enc_lambda_layers functions to try:
K.softsign
a = 1
longtanh = lambda x: K.tanh(x) * K.log(2 + a * abs(x))
"""
# naming
model_name = name
# volume size data
ndims = len(input_shape) - 1
if isinstance(pool_size, int):
pool_size = (pool_size,) * ndims
# get encoding model
enc_model = conv_enc(nb_features,
input_shape,
nb_levels,
conv_size,
name=model_name,
feat_mult=feat_mult,
pool_size=pool_size,
padding=padding,
activation=activation,
use_residuals=use_residuals,
nb_conv_per_level=nb_conv_per_level,
conv_dropout=conv_dropout,
batch_norm=batch_norm)
# middle AE structure
if single_model:
in_input_shape = None
in_model = enc_model
else:
in_input_shape = enc_model.output.shape.as_list()[1:]
in_model = None
mid_ae_model = single_ae(enc_size,
in_input_shape,
conv_size=conv_size,
name=model_name,
ae_type=ae_type,
input_model=in_model,
batch_norm=enc_batch_norm,
enc_lambda_layers=enc_lambda_layers,
include_mu_shift_layer=include_mu_shift_layer,
do_vae=do_vae)
# decoder
if single_model:
in_input_shape = None
in_model = mid_ae_model
else:
in_input_shape = mid_ae_model.output.shape.as_list()[1:]
in_model = None
dec_model = conv_dec(nb_features,
in_input_shape,
nb_levels,
conv_size,
nb_labels,
name=model_name,
feat_mult=feat_mult,
pool_size=pool_size,
use_skip_connections=False,
padding=padding,
activation=activation,
use_residuals=use_residuals,
final_pred_activation='linear',
nb_conv_per_level=nb_conv_per_level,
batch_norm=batch_norm,
conv_dropout=conv_dropout,
input_model=in_model)
if add_prior_layer:
dec_model = add_prior(dec_model,
[*input_shape[:-1],nb_labels],
name=model_name,
prefix=model_name + '_prior',
use_logp=use_logp,
final_pred_activation=final_pred_activation,
add_prior_layer_reg=add_prior_layer_reg)
if single_model:
return dec_model
else:
return (dec_model, mid_ae_model, enc_model)
def conv_enc(nb_features,
input_shape,
nb_levels,
conv_size,
name=None,
prefix=None,
feat_mult=1,
pool_size=2,
dilation_rate_mult=1,
padding='same',
activation='elu',
layer_nb_feats=None,
use_residuals=False,
nb_conv_per_level=2,
conv_dropout=0,
batch_norm=None):
"""
Fully Convolutional Encoder
"""
# naming
model_name = name
if prefix is None:
prefix = model_name
# volume size data
ndims = len(input_shape) - 1
input_shape = tuple(input_shape)
if isinstance(pool_size, int):
pool_size = (pool_size,) * ndims
# prepare layers
convL = getattr(KL, 'Conv%dD' % ndims)
conv_kwargs = {'padding': padding, 'activation': activation}
maxpool = getattr(KL, 'MaxPooling%dD' % ndims)
# first layer: input
name = '%s_input' % prefix
last_tensor = KL.Input(shape=input_shape, name=name)
input_tensor = last_tensor
# down arm:
# add nb_levels of conv + ReLu + conv + ReLu. Pool after each of first nb_levels - 1 layers
lfidx = 0
for level in range(nb_levels):
lvl_first_tensor = last_tensor
nb_lvl_feats = np.round(nb_features*feat_mult**level).astype(int)
conv_kwargs['dilation_rate'] = dilation_rate_mult**level
for conv in range(nb_conv_per_level):
if layer_nb_feats is not None:
nb_lvl_feats = layer_nb_feats[lfidx]
lfidx += 1
name = '%s_conv_downarm_%d_%d' % (prefix, level, conv)
if conv < (nb_conv_per_level-1) or (not use_residuals):
last_tensor = convL(nb_lvl_feats, conv_size, **conv_kwargs, name=name)(last_tensor)
else: # no activation
last_tensor = convL(nb_lvl_feats, conv_size, padding=padding, name=name)(last_tensor)
if conv_dropout > 0:
# conv dropout along feature space only
name = '%s_dropout_downarm_%d_%d' % (prefix, level, conv)
noise_shape = [None, *[1]*ndims, nb_lvl_feats]
last_tensor = KL.Dropout(conv_dropout, noise_shape=noise_shape)(last_tensor)
if use_residuals:
convarm_layer = last_tensor
# the "add" layer is the original input
# However, it may not have the right number of features to be added
nb_feats_in = lvl_first_tensor.get_shape()[-1]
nb_feats_out = convarm_layer.get_shape()[-1]
add_layer = lvl_first_tensor
if nb_feats_in > 1 and nb_feats_out > 1 and (nb_feats_in != nb_feats_out):
name = '%s_expand_down_merge_%d' % (prefix, level)
last_tensor = convL(nb_lvl_feats, conv_size, **conv_kwargs, name=name)(lvl_first_tensor)
add_layer = last_tensor
if conv_dropout > 0:
name = '%s_dropout_down_merge_%d_%d' % (prefix, level, conv)
noise_shape = [None, *[1]*ndims, nb_lvl_feats]
last_tensor = KL.Dropout(conv_dropout, noise_shape=noise_shape)(last_tensor)
name = '%s_res_down_merge_%d' % (prefix, level)
last_tensor = KL.add([add_layer, convarm_layer], name=name)
name = '%s_res_down_merge_act_%d' % (prefix, level)
last_tensor = KL.Activation(activation, name=name)(last_tensor)
if batch_norm is not None:
name = '%s_bn_down_%d' % (prefix, level)
last_tensor = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# max pool if we're not at the last level
if level < (nb_levels - 1):
name = '%s_maxpool_%d' % (prefix, level)
last_tensor = maxpool(pool_size=pool_size, name=name, padding=padding)(last_tensor)
# create the model and return
model = Model(inputs=input_tensor, outputs=[last_tensor], name=model_name)
return model
def conv_dec(nb_features,
input_shape,
nb_levels,
conv_size,
nb_labels,
name=None,
prefix=None,
feat_mult=1,
pool_size=2,
use_skip_connections=False,
padding='same',
dilation_rate_mult=1,
activation='elu',
use_residuals=False,
final_pred_activation='softmax',
nb_conv_per_level=2,
layer_nb_feats=None,
batch_norm=None,
conv_dropout=0,
input_model=None):
"""
Fully Convolutional Decoder
Parameters:
...
use_skip_connections (bool): if true, turns an Enc-Dec to a U-Net.
If true, input_tensor and tensors are required.
It assumes a particular naming of layers. conv_enc...
"""
# naming
model_name = name
if prefix is None:
prefix = model_name
# if using skip connections, make sure need to use them.
if use_skip_connections:
assert input_model is not None, "is using skip connections, tensors dictionary is required"
# first layer: input
input_name = '%s_input' % prefix
if input_model is None:
input_tensor = KL.Input(shape=input_shape, name=input_name)
last_tensor = input_tensor
else:
input_tensor = input_model.input
last_tensor = input_model.output
input_shape = last_tensor.shape.as_list()[1:]
# vol size info
ndims = len(input_shape) - 1
input_shape = tuple(input_shape)
if isinstance(pool_size, int):
pool_size = (pool_size,) * ndims
# prepare layers
convL = getattr(KL, 'Conv%dD' % ndims)
conv_kwargs = {'padding': padding, 'activation': activation}
upsample = getattr(KL, 'UpSampling%dD' % ndims)
# up arm:
# nb_levels - 1 layers of Deconvolution3D
# (approx via up + conv + ReLu) + merge + conv + ReLu + conv + ReLu
lfidx = 0
for level in range(nb_levels - 1):
nb_lvl_feats = np.round(nb_features*feat_mult**(nb_levels-2-level)).astype(int)
conv_kwargs['dilation_rate'] = dilation_rate_mult**(nb_levels-2-level)
# upsample matching the max pooling layers size
name = '%s_up_%d' % (prefix, nb_levels + level)
last_tensor = upsample(size=pool_size, name=name)(last_tensor)
up_tensor = last_tensor
# merge layers combining previous layer
# TODO: add Cropping3D or Cropping2D if 'valid' padding
if use_skip_connections:
conv_name = '%s_conv_downarm_%d_%d' % (prefix, nb_levels - 2 - level, nb_conv_per_level - 1)
cat_tensor = input_model.get_layer(conv_name).output
name = '%s_merge_%d' % (prefix, nb_levels + level)
last_tensor = KL.concatenate([cat_tensor, last_tensor], axis=ndims+1, name=name)
# convolution layers
for conv in range(nb_conv_per_level):
if layer_nb_feats is not None:
nb_lvl_feats = layer_nb_feats[lfidx]
lfidx += 1
name = '%s_conv_uparm_%d_%d' % (prefix, nb_levels + level, conv)
if conv < (nb_conv_per_level-1) or (not use_residuals):
last_tensor = convL(nb_lvl_feats, conv_size, **conv_kwargs, name=name)(last_tensor)
else:
last_tensor = convL(nb_lvl_feats, conv_size, padding=padding, name=name)(last_tensor)
if conv_dropout > 0:
name = '%s_dropout_uparm_%d_%d' % (prefix, level, conv)
noise_shape = [None, *[1]*ndims, nb_lvl_feats]
last_tensor = KL.Dropout(conv_dropout, noise_shape=noise_shape)(last_tensor)
# residual block
if use_residuals:
# the "add" layer is the original input
# However, it may not have the right number of features to be added
add_layer = up_tensor
nb_feats_in = add_layer.get_shape()[-1]
nb_feats_out = last_tensor.get_shape()[-1]
if nb_feats_in > 1 and nb_feats_out > 1 and (nb_feats_in != nb_feats_out):
name = '%s_expand_up_merge_%d' % (prefix, level)
add_layer = convL(nb_lvl_feats, conv_size, **conv_kwargs, name=name)(add_layer)
if conv_dropout > 0:
name = '%s_dropout_up_merge_%d_%d' % (prefix, level, conv)
noise_shape = [None, *[1]*ndims, nb_lvl_feats]
last_tensor = KL.Dropout(conv_dropout, noise_shape=noise_shape)(last_tensor)
name = '%s_res_up_merge_%d' % (prefix, level)
last_tensor = KL.add([last_tensor, add_layer], name=name)
name = '%s_res_up_merge_act_%d' % (prefix, level)
last_tensor = KL.Activation(activation, name=name)(last_tensor)
if batch_norm is not None:
name = '%s_bn_up_%d' % (prefix, level)
last_tensor = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# Compute likelyhood prediction (no activation yet)
name = '%s_likelihood' % prefix
last_tensor = convL(nb_labels, 1, activation=None, name=name)(last_tensor)
like_tensor = last_tensor
# output prediction layer
# we use a softmax to compute P(L_x|I) where x is each location
if final_pred_activation == 'softmax':
print("using final_pred_activation %s for %s" % (final_pred_activation, model_name))
name = '%s_prediction' % prefix
softmax_lambda_fcn = lambda x: keras.activations.softmax(x, axis=ndims + 1)
pred_tensor = KL.Lambda(softmax_lambda_fcn, name=name)(last_tensor)
# otherwise create a layer that does nothing.
else:
name = '%s_prediction' % prefix
pred_tensor = KL.Activation('linear', name=name)(like_tensor)
# create the model and retun
model = Model(inputs=input_tensor, outputs=pred_tensor, name=model_name)
return model
def add_prior(input_model,
prior_shape,
name='prior_model',
prefix=None,
use_logp=True,
final_pred_activation='softmax',
add_prior_layer_reg=0):
"""
Append post-prior layer to a given model
"""
# naming
model_name = name
if prefix is None:
prefix = model_name
# prior input layer
prior_input_name = '%s-input' % prefix
prior_tensor = KL.Input(shape=prior_shape, name=prior_input_name)
prior_tensor_input = prior_tensor
like_tensor = input_model.output
# operation varies depending on whether we log() prior or not.
if use_logp:
# name = '%s-log' % prefix
# prior_tensor = KL.Lambda(_log_layer_wrap(add_prior_layer_reg), name=name)(prior_tensor)
print("Breaking change: use_logp option now requires log input!", file=sys.stderr)
merge_op = KL.add
else:
# using sigmoid to get the likelihood values between 0 and 1
# note: they won't add up to 1.
name = '%s_likelihood_sigmoid' % prefix
like_tensor = KL.Activation('sigmoid', name=name)(like_tensor)
merge_op = KL.multiply
# merge the likelihood and prior layers into posterior layer
name = '%s_posterior' % prefix
post_tensor = merge_op([prior_tensor, like_tensor], name=name)
# output prediction layer
# we use a softmax to compute P(L_x|I) where x is each location
pred_name = '%s_prediction' % prefix
if final_pred_activation == 'softmax':
assert use_logp, 'cannot do softmax when adding prior via P()'
print("using final_pred_activation %s for %s" % (final_pred_activation, model_name))
softmax_lambda_fcn = lambda x: keras.activations.softmax(x, axis=-1)
pred_tensor = KL.Lambda(softmax_lambda_fcn, name=pred_name)(post_tensor)
else:
pred_tensor = KL.Activation('linear', name=pred_name)(post_tensor)
# create the model
model_inputs = [*input_model.inputs, prior_tensor_input]
model = Model(inputs=model_inputs, outputs=[pred_tensor], name=model_name)
# compile
return model
def single_ae(enc_size,
input_shape,
name='single_ae',
prefix=None,
ae_type='dense', # 'dense', or 'conv'
conv_size=None,
input_model=None,
enc_lambda_layers=None,
batch_norm=True,
padding='same',
activation=None,
include_mu_shift_layer=False,
do_vae=False):
"""single-layer Autoencoder (i.e. input - encoding - output"""
# naming
model_name = name
if prefix is None:
prefix = model_name
if enc_lambda_layers is None:
enc_lambda_layers = []
# prepare input
input_name = '%s_input' % prefix
if input_model is None:
assert input_shape is not None, 'input_shape of input_model is necessary'
input_tensor = KL.Input(shape=input_shape, name=input_name)
last_tensor = input_tensor
else:
input_tensor = input_model.input
last_tensor = input_model.output
input_shape = last_tensor.shape.as_list()[1:]
input_nb_feats = last_tensor.shape.as_list()[-1]
# prepare conv type based on input
if ae_type == 'conv':
ndims = len(input_shape) - 1
convL = getattr(KL, 'Conv%dD' % ndims)
assert conv_size is not None, 'with conv ae, need conv_size'
conv_kwargs = {'padding': padding, 'activation': activation}
# if want to go through a dense layer in the middle of the U, need to:
# - flatten last layer if not flat
# - do dense encoding and decoding
# - unflatten (rehsape spatially) at end
if ae_type == 'dense' and len(input_shape) > 1:
name = '%s_ae_%s_down_flat' % (prefix, ae_type)
last_tensor = KL.Flatten(name=name)(last_tensor)
# recall this layer
pre_enc_layer = last_tensor
# encoding layer
if ae_type == 'dense':
assert len(enc_size) == 1, "enc_size should be of length 1 for dense layer"
enc_size_str = ''.join(['%d_' % d for d in enc_size])[:-1]
name = '%s_ae_mu_enc_dense_%s' % (prefix, enc_size_str)
last_tensor = KL.Dense(enc_size[0], name=name)(pre_enc_layer)
else: # convolution
# convolve then resize. enc_size should be [nb_dim1, nb_dim2, ..., nb_feats]
assert len(enc_size) == len(input_shape), \
"encoding size does not match input shape %d %d" % (len(enc_size), len(input_shape))
if list(enc_size)[:-1] != list(input_shape)[:-1] and \
all([f is not None for f in input_shape[:-1]]) and \
all([f is not None for f in enc_size[:-1]]):
assert len(enc_size) - 1 == 2, "Sorry, I have not yet implemented non-2D resizing -- need to check out interpn!"
name = '%s_ae_mu_enc_conv' % (prefix)
last_tensor = convL(enc_size[-1], conv_size, name=name, **conv_kwargs)(pre_enc_layer)
name = '%s_ae_mu_enc' % (prefix)
resize_fn = lambda x: tf.image.resize_bilinear(x, enc_size[:-1])
last_tensor = KL.Lambda(resize_fn, name=name)(last_tensor)
elif enc_size[-1] is None: # convolutional, but won't tell us bottleneck
name = '%s_ae_mu_enc' % (prefix)
last_tensor = KL.Lambda(lambda x: x, name=name)(pre_enc_layer)
else:
name = '%s_ae_mu_enc' % (prefix)
last_tensor = convL(enc_size[-1], conv_size, name=name, **conv_kwargs)(pre_enc_layer)
if include_mu_shift_layer:
# shift
name = '%s_ae_mu_shift' % (prefix)
last_tensor = layers.LocalBias(name=name)(last_tensor)
# encoding clean-up layers
for layer_fcn in enc_lambda_layers:
lambda_name = layer_fcn.__name__
name = '%s_ae_mu_%s' % (prefix, lambda_name)
last_tensor = KL.Lambda(layer_fcn, name=name)(last_tensor)
if batch_norm is not None:
name = '%s_ae_mu_bn' % (prefix)
last_tensor = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# have a simple layer that does nothing to have a clear name before sampling
name = '%s_ae_mu' % (prefix)
last_tensor = KL.Lambda(lambda x: x, name=name)(last_tensor)
# if doing variational AE, will need the sigma layer as well.
if do_vae:
mu_tensor = last_tensor
# encoding layer
if ae_type == 'dense':
name = '%s_ae_sigma_enc_dense_%s' % (prefix, enc_size_str)
last_tensor = KL.Dense(enc_size[0], name=name)(pre_enc_layer)
else:
if list(enc_size)[:-1] != list(input_shape)[:-1] and \
all([f is not None for f in input_shape[:-1]]) and \
all([f is not None for f in enc_size[:-1]]):
assert len(enc_size) - 1 == 2, "Sorry, I have not yet implemented non-2D resizing..."
name = '%s_ae_sigma_enc_conv' % (prefix)
last_tensor = convL(enc_size[-1], conv_size, name=name, **conv_kwargs)(pre_enc_layer)
name = '%s_ae_sigma_enc' % (prefix)
resize_fn = lambda x: tf.image.resize_bilinear(x, enc_size[:-1])
last_tensor = KL.Lambda(resize_fn, name=name)(last_tensor)
elif enc_size[-1] is None: # convolutional, but won't tell us bottleneck
name = '%s_ae_sigma_enc' % (prefix)
last_tensor = convL(pre_enc_layer.shape.as_list()[-1], conv_size, name=name, **conv_kwargs)(pre_enc_layer)
# cannot use lambda, then mu and sigma will be same layer.
# last_tensor = KL.Lambda(lambda x: x, name=name)(pre_enc_layer)
else:
name = '%s_ae_sigma_enc' % (prefix)
last_tensor = convL(enc_size[-1], conv_size, name=name, **conv_kwargs)(pre_enc_layer)
# encoding clean-up layers
for layer_fcn in enc_lambda_layers:
lambda_name = layer_fcn.__name__
name = '%s_ae_sigma_%s' % (prefix, lambda_name)
last_tensor = KL.Lambda(layer_fcn, name=name)(last_tensor)
if batch_norm is not None:
name = '%s_ae_sigma_bn' % (prefix)
last_tensor = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# have a simple layer that does nothing to have a clear name before sampling
name = '%s_ae_sigma' % (prefix)
last_tensor = KL.Lambda(lambda x: x, name=name)(last_tensor)
logvar_tensor = last_tensor
# VAE sampling
sampler = _VAESample().sample_z
name = '%s_ae_sample' % (prefix)
last_tensor = KL.Lambda(sampler, name=name)([mu_tensor, logvar_tensor])
if include_mu_shift_layer:
# shift
name = '%s_ae_sample_shift' % (prefix)
last_tensor = layers.LocalBias(name=name)(last_tensor)
# decoding layer
if ae_type == 'dense':
name = '%s_ae_%s_dec_flat_%s' % (prefix, ae_type, enc_size_str)
last_tensor = KL.Dense(np.prod(input_shape), name=name)(last_tensor)
# unflatten if dense method
if len(input_shape) > 1:
name = '%s_ae_%s_dec' % (prefix, ae_type)
last_tensor = KL.Reshape(input_shape, name=name)(last_tensor)
else:
if list(enc_size)[:-1] != list(input_shape)[:-1] and \
all([f is not None for f in input_shape[:-1]]) and \
all([f is not None for f in enc_size[:-1]]):
name = '%s_ae_mu_dec' % (prefix)
resize_fn = lambda x: tf.image.resize_bilinear(x, input_shape[:-1])
last_tensor = KL.Lambda(resize_fn, name=name)(last_tensor)
name = '%s_ae_%s_dec' % (prefix, ae_type)
last_tensor = convL(input_nb_feats, conv_size, name=name, **conv_kwargs)(last_tensor)
if batch_norm is not None:
name = '%s_bn_ae_%s_dec' % (prefix, ae_type)
last_tensor = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# create the model and retun
model = Model(inputs=input_tensor, outputs=[last_tensor], name=model_name)
return model
def design_dnn(nb_features, input_shape, nb_levels, conv_size, nb_labels,
feat_mult=1,
pool_size=2,
padding='same',
activation='elu',
final_layer='dense-sigmoid',
conv_dropout=0,
conv_maxnorm=0,
nb_input_features=1,
batch_norm=False,
name=None,
prefix=None,
use_strided_convolution_maxpool=True,
nb_conv_per_level=2):
"""
"deep" cnn with dense or global max pooling layer @ end...
Could use sequential...
"""
model_name = name
if model_name is None:
model_name = 'model_1'
if prefix is None:
prefix = model_name
ndims = len(input_shape)
input_shape = tuple(input_shape)
convL = getattr(KL, 'Conv%dD' % ndims)
maxpool = KL.MaxPooling3D if len(input_shape) == 3 else KL.MaxPooling2D
if isinstance(pool_size, int):
pool_size = (pool_size,) * ndims
# kwargs for the convolution layer
conv_kwargs = {'padding': padding, 'activation': activation}
if conv_maxnorm > 0:
conv_kwargs['kernel_constraint'] = maxnorm(conv_maxnorm)
# initialize a dictionary
enc_tensors = {}
# first layer: input
name = '%s_input' % prefix
enc_tensors[name] = KL.Input(shape=input_shape + (nb_input_features,), name=name)
last_tensor = enc_tensors[name]
# down arm:
# add nb_levels of conv + ReLu + conv + ReLu. Pool after each of first nb_levels - 1 layers
for level in range(nb_levels):
for conv in range(nb_conv_per_level):
if conv_dropout > 0:
name = '%s_dropout_%d_%d' % (prefix, level, conv)
enc_tensors[name] = KL.Dropout(conv_dropout)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_conv_%d_%d' % (prefix, level, conv)
nb_lvl_feats = np.round(nb_features*feat_mult**level).astype(int)
enc_tensors[name] = convL(nb_lvl_feats, conv_size, **conv_kwargs, name=name)(last_tensor)
last_tensor = enc_tensors[name]
# max pool
if use_strided_convolution_maxpool:
name = '%s_strided_conv_%d' % (prefix, level)
enc_tensors[name] = convL(nb_lvl_feats, pool_size, **conv_kwargs, name=name)(last_tensor)
last_tensor = enc_tensors[name]
else:
name = '%s_maxpool_%d' % (prefix, level)
enc_tensors[name] = maxpool(pool_size=pool_size, name=name, padding=padding)(last_tensor)
last_tensor = enc_tensors[name]
# dense layer
if final_layer == 'dense-sigmoid':
name = "%s_flatten" % prefix
enc_tensors[name] = KL.Flatten(name=name)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_dense' % prefix
enc_tensors[name] = KL.Dense(1, name=name, activation="sigmoid")(last_tensor)
elif final_layer == 'dense-tanh':
name = "%s_flatten" % prefix
enc_tensors[name] = KL.Flatten(name=name)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_dense' % prefix
enc_tensors[name] = KL.Dense(1, name=name)(last_tensor)
last_tensor = enc_tensors[name]
# Omittting BatchNorm for now, it seems to have a cpu vs gpu problem
# https://github.com/tensorflow/tensorflow/pull/8906
# https://github.com/fchollet/keras/issues/5802
# name = '%s_%s_bn' % prefix
# enc_tensors[name] = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
# last_tensor = enc_tensors[name]
name = '%s_%s_tanh' % prefix
enc_tensors[name] = KL.Activation(activation="tanh", name=name)(last_tensor)
elif final_layer == 'dense-softmax':
name = "%s_flatten" % prefix
enc_tensors[name] = KL.Flatten(name=name)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_dense' % prefix
enc_tensors[name] = KL.Dense(nb_labels, name=name, activation="softmax")(last_tensor)
# global max pooling layer
elif final_layer == 'myglobalmaxpooling':
name = '%s_batch_norm' % prefix
enc_tensors[name] = KL.BatchNormalization(axis=batch_norm, name=name)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_global_max_pool' % prefix
enc_tensors[name] = KL.Lambda(_global_max_nd, name=name)(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_global_max_pool_reshape' % prefix
enc_tensors[name] = KL.Reshape((1, 1), name=name)(last_tensor)
last_tensor = enc_tensors[name]
# cannot do activation in lambda layer. Could code inside, but will do extra lyaer
name = '%s_global_max_pool_sigmoid' % prefix
enc_tensors[name] = KL.Conv1D(1, 1, name=name, activation="sigmoid", use_bias=True)(last_tensor)
elif final_layer == 'globalmaxpooling':
name = '%s_conv_to_featmaps' % prefix
enc_tensors[name] = KL.Conv3D(2, 1, name=name, activation="relu")(last_tensor)
last_tensor = enc_tensors[name]
name = '%s_global_max_pool' % prefix
enc_tensors[name] = KL.GlobalMaxPooling3D(name=name)(last_tensor)
last_tensor = enc_tensors[name]
# cannot do activation in lambda layer. Could code inside, but will do extra lyaer
name = '%s_global_max_pool_softmax' % prefix
enc_tensors[name] = KL.Activation('softmax', name=name)(last_tensor)
last_tensor = enc_tensors[name]
# create the model
model = Model(inputs=[enc_tensors['%s_input' % prefix]], outputs=[last_tensor], name=model_name)
return model
###############################################################################
# Helper function
###############################################################################
def _global_max_nd(xtens):
ytens = K.batch_flatten(xtens)
return K.max(ytens, 1, keepdims=True)
def _log_layer_wrap(reg=K.epsilon()):
def _log_layer(tens):
return K.log(tens + reg)
return _log_layer
# def _global_max_nd(x):
# return K.exp(x)
class _VAESample():
def __init__(self): #, nb_z):
# self.nb_z = nb_z
pass
def sample_z(self, args):
mu, log_var = args
# shape = (K.shape(mu)[0], self.nb_z)
shape = K.shape(mu)
eps = K.random_normal(shape=shape, mean=0., stddev=1.)
return mu + K.exp(log_var / 2) * eps
def _softmax(x, axis=-1, alpha=1):
"""
building on keras implementation, allow alpha parameter
Softmax activation function.
# Arguments
x : Tensor.
axis: Integer, axis along which the softmax normalization is applied.
alpha: a value to multiply all x
# Returns
Tensor, output of softmax transformation.
# Raises
ValueError: In case `dim(x) == 1`.
"""
x = alpha * x
ndim = K.ndim(x)
if ndim == 2:
return K.softmax(x)
elif ndim > 2:
e = K.exp(x - K.max(x, axis=axis, keepdims=True))
s = K.sum(e, axis=axis, keepdims=True)
return e / s
else:
raise ValueError('Cannot apply softmax to a tensor that is 1D')

models.py at commit ceab402, under MIT · at the source

Overview

Authors: Henrik Skibbe^1,2, Michal Byra³, Akiya Watakabe⁴, Tetsuo Yamamori⁵, Marco Reisert⁶

Department of Informatics,Faculty of Informatics, Matsuyama University, Matsuyama, Ehime Japan
Brain Image Analysis Unit, RIKEN Center for Brain Science, Wako, Saitama Japan
Institute of Fundamental Technological Research, Polish Academy of Sciences, Warsaw, Poland
Laboratory for Molecular Mechanisms of Brain Development, RIKEN Center for Brain Science, Wako, Saitama Japan
Department of Biological Functions and Regulation, CIEM (Central Institute of Experimental Medicine and Life Science), Kawasaki, Japan
Department of Radiology, Medical Physics, and Department of Stereotactic and Functional Neurosurgery, Medical Center Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany

Institutions: RIKEN Center for Brain Science (Japan); Matsuyama University (Japan); Institute of Fundamental Technological Research (Poland); Polish Academy of Sciences (Poland); University of Freiburg (Germany); University Medical Center Freiburg (Germany)

Journal: Scientific reports, volume 16, issue 1, article 14386

Dates: received 30 April 2025; accepted 16 March 2026; published online 21 March 2026

Type: Research article · Language: English

License: CC BY-NC-ND

Identifiers: DOI 10.1038/s41598-026-44858-x · PMID 41865161 · PMCID PMC13144365 · OpenAlex W7139991225

Open access: gold, a free copy (OpenAlex)

Status: code verified (how the code is found and verified)

Categories: structural MRI / diffusion (modality), human (organism)

Methods: Machine learning

Keywords: Engineering, Mathematics and computing

MeSH: Brain*, Image Processing, Computer-Assisted*, Imaging, Three-Dimensional*, Algorithms, Animals, Convolutional Neural Networks, Humans, Magnetic Resonance Imaging (* major topic)

Topic: Advanced Neural Network Applications (Computer Vision and Pattern Recognition, Computer Science), according to OpenAlex

Funding: Japan Agency for Medical Research and Development (AMED) (JP23wm0625001)

Citations: not cited yet (Europe PMC); 35 references in the paper

Abstract

The abstract is not reproduced here: the paper's license (CC BY-NC-ND) does not allow it. Read it in the paper, at the publisher or on Europe PMC.

Repositories

Its files are read in the Code ↔ Paper reader above, with 2 matches between paragraphs and lines of code.

cwmok/LapIRN

License: MIT

State: the link answers, verified on 30 September 2026

Evidence: files inventoried

Commit: c17fb7564ca752d37e97baf12cafff44e01da668, 20 July 2023

Languages: Python (11)

Size: 25 files, 11 scripts

Software Heritage: not archived

Found in: the text, “Reference approaches”

Holds: README, license file

Not found: CITATION.cff, environment file, tests, continuous integration, documentation

Tools: NumPy (11 files), PyTorch (11 files), NiBabel (2 files)

Availability: 1 check, the latest on 30 September 2026: the link answers

30 September 2026: the link answers

13 files

Code/Functions.py, Python, 213 lines
Code/Test_LapIRN_diff.py, Python, 92 lines
Code/Test_LapIRN_diff_resize.py, Python, 100 lines
Code/Test_LapIRN_disp.py, Python, 85 lines
Code/Test_LapIRN_disp_resize_unorm_beta.py, Python, 110 lines
Code/Train_LapIRN_diff.py, Python, 391 lines
Code/Train_LapIRN_diff_resize.py, Python, 471 lines
Code/Train_LapIRN_disp.py, Python, 391 lines
Code/Train_LapIRN_disp_resize_unorm_beta.py, Python, 470 lines
Code/count_parameters.py, Python, 53 lines
Code/miccai2020_model_stage.py, Python, 1,280 lines
LICENSE.md, License, 22 lines
README.md, Text, 56 lines

At the source: github.com/cwmok/LapIRN

MungoMeng/Registration-NICE-Net

License: MIT

State: the link answers, verified on 30 September 2026

Evidence: files inventoried

Commit: ceab4022123a9e50524c2d72f184959807fd7039, 18 September 2023

Languages: Python (26)

Size: 29 files, 26 scripts

Software Heritage: not archived

Found in: the text, “Reference approaches”

Holds: README, license file

Not found: CITATION.cff, environment file, tests, continuous integration, documentation

Tools: NumPy (20 files), Keras (11 files), TensorFlow (8 files), Matplotlib (5 files), SciPy (5 files), NiBabel (3 files), NEURON (2 files), Pillow (1 file)

Availability: 1 check, the latest on 30 September 2026: the link answers

30 September 2026: the link answers

28 files

NICE-Net/datagenerators.py, Python, 113 lines
NICE-Net/ext/medipy/__init__.py, Python, 1 line
NICE-Net/ext/medipy/metrics.py, Python, 52 lines
NICE-Net/ext/neuron/__init__.py, Python, 10 lines
NICE-Net/ext/neuron/callbacks.py, Python, 626 lines
NICE-Net/ext/neuron/dataproc.py, Python, 420 lines
NICE-Net/ext/neuron/generators.py, Python, 1,415 lines
NICE-Net/ext/neuron/inits.py, Python, 19 lines
NICE-Net/ext/neuron/layers.py, Python, 853 lines
NICE-Net/ext/neuron/metrics.py, Python, 435 lines
NICE-Net/ext/neuron/models.py, Python, 1,009 lines, 1 match
NICE-Net/ext/neuron/plot.py, Python, 254 lines
NICE-Net/ext/neuron/utils.py, Python, 1,287 lines
NICE-Net/ext/pynd/__init__.py, Python, 2 lines
NICE-Net/ext/pynd/imutils.py, Python, 11 lines
NICE-Net/ext/pynd/ndutils.py, Python, 436 lines
NICE-Net/ext/pynd/segutils.py, Python, 122 lines
NICE-Net/ext/pytoolslib/__init__.py, Python, 4 lines
NICE-Net/ext/pytoolslib/iniparse.py, Python, 211 lines
NICE-Net/ext/pytoolslib/patchlib.py, Python, 451 lines
NICE-Net/ext/pytoolslib/plotting.py, Python, 63 lines
NICE-Net/ext/pytoolslib/timer.py, Python, 31 lines
NICE-Net/losses.py, Python, 117 lines
NICE-Net/networks.py, Python, 493 lines
NICE-Net/test.py, Python, 131 lines
NICE-Net/train.py, Python, 136 lines
LICENSE, License, 21 lines
README.md, Text, 11 lines

At the source: github.com/MungoMeng/Registration-NICE-Net

MungoMeng/Registration-NICE-Trans

License: GPL-3.0

State: the link answers, verified on 30 September 2026

Evidence: files inventoried

Commit: f213adeac0868ac5631e354ff0357e497b8ff78b, 15 March 2024

Languages: Python (5)

Size: 8 files, 5 scripts

Software Heritage: not archived

Found in: the text, “Reference approaches”

Holds: README, license file

Not found: CITATION.cff, environment file, tests, continuous integration, documentation

Tools: NumPy (5 files), PyTorch (4 files), SciPy (3 files), NiBabel (1 file)

Availability: 1 check, the latest on 30 September 2026: the link answers

30 September 2026: the link answers

7 files

NICE-Trans/datagenerators.py, Python, 91 lines
NICE-Trans/losses.py, Python, 118 lines
NICE-Trans/networks.py, Python, 710 lines
NICE-Trans/test.py, Python, 156 lines
NICE-Trans/train.py, Python, 229 lines
LICENSE, License, 674 lines
README.md, Text, 10 lines

At the source: github.com/MungoMeng/Registration-NICE-Trans

BrainImageAnalysis/INRsRegExp

License: MIT

State: the link answers, verified on 31 October 2026

Evidence: files inventoried

Commit: 74036e2b761b252375dfda40dca3fd6bfa19deaa, 5 December 2023

Languages: Python (5), Jupyter (2)

Size: 16 files, 7 scripts

Software Heritage: not archived

Found in: the text, “Reference approaches”

Holds: README, license file, 2 notebooks

Not found: CITATION.cff, environment file, tests, continuous integration, documentation

Tools: NumPy (6 files), PyTorch (6 files), Matplotlib (3 files), NiBabel (1 file), SciPy (1 file)

Availability: 2 checks, the latest on 31 October 2026: the link answers

31 October 2026: the link answers
27 September 2026: the link answers

9 files

example_custom.ipynb, Jupyter, 240 lines
example_functions.ipynb, Jupyter, 96 lines
libs/__init__.py, Python, 1 line
libs/codes.py, Python, 1,693 lines, 1 match
libs/data.py, Python, 325 lines
libs/losses.py, Python, 190 lines
libs/networks.py, Python, 1,578 lines
LICENSE, License, 21 lines
README.md, Text, 31 lines

At the source: github.com/BrainImageAnalysis/INRsRegExp

BrainImageAnalysis/PatchMorph

License: Apache-2.0

State: the link answers, verified on 30 September 2026

Evidence: files inventoried

Commit: 562eeac00382cad3609ccf115223308ef7268819, 19 April 2024

Size: 1 file, 0 scripts

Software Heritage: not archived

Found in: the text, “Conclusion”

Holds: license file

Not found: README, CITATION.cff, environment file, tests, continuous integration, documentation

Availability: 1 check, the latest on 30 September 2026: the link answers

30 September 2026: the link answers

1 file

LICENSE, License, 201 lines

At the source: github.com/BrainImageAnalysis/PatchMorph

The paper's code and data availability statement is in the Data section.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

5 repositories of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
49 scripts, each with its path and the digest of its content;
2 matches between paragraphs of the paper and lines of the code (method lexical-v1);
neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

Datasets cited

zenodo:19041876, at Zenodo; found in “Data availability”

Code and data availability statement

The paper has a code and data availability statement. Its license (CC BY-NC-ND) does not allow reproducing it here; in short, from what the harvester recognized in it:

it points to a dataset: Zenodo 19041876

Read it in the paper: doi.org/10.1038/s41598-026-44858-x.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 30 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 5 authors, 2 keywords, 8 MeSH terms, 1 funder, 18 references.

Cite

This paper

Skibbe, H., Byra, M., Watakabe, A., Yamamori, T., & Reisert, M. (2026). Memory efficient training for 3D brain image registration networks using PatchMorph. Scientific reports, 16(1), 14386. https://doi.org/10.1038/s41598-026-44858-x

BibTeX

@article{skibbe2026memory, author = {Skibbe, Henrik and Byra, Michal and Watakabe, Akiya and Yamamori, Tetsuo and Reisert, Marco}, title = {{Memory efficient training for 3D brain image registration networks using PatchMorph}}, journal = {Scientific reports}, year = {2026}, month = mar, volume = {16}, number = {1}, pages = {14386}, publisher = {Nature Publishing Group}, issn = {2045-2322}, doi = {10.1038/s41598-026-44858-x}, url = {https://doi.org/10.1038/s41598-026-44858-x}, pmid = {41865161}, pmcid = {PMC13144365} }

RIS

TY - JOUR AU - Skibbe, Henrik AU - Byra, Michal AU - Watakabe, Akiya AU - Yamamori, Tetsuo AU - Reisert, Marco TI - Memory efficient training for 3D brain image registration networks using PatchMorph T2 - Scientific reports J2 - Sci Rep PY - 2026 DA - 2026/03/21 VL - 16 IS - 1 SP - 14386 SN - 2045-2322 PB - Nature Publishing Group DO - 10.1038/s41598-026-44858-x UR - https://doi.org/10.1038/s41598-026-44858-x LA - en ER -

CSL-JSON

{ "id": "10.1038/s41598-026-44858-x", "type": "article-journal", "title": "Memory efficient training for 3D brain image registration networks using PatchMorph", "container-title": "Scientific reports", "author": [ { "family": "Skibbe", "given": "Henrik" }, { "family": "Byra", "given": "Michal" }, { "family": "Watakabe", "given": "Akiya" }, { "family": "Yamamori", "given": "Tetsuo" }, { "family": "Reisert", "given": "Marco" } ], "container-title-short": "Sci Rep", "volume": "16", "issue": "1", "page": "14386", "DOI": "10.1038/s41598-026-44858-x", "PMID": "41865161", "PMCID": "PMC13144365", "ISSN": "2045-2322", "publisher": "Nature Publishing Group", "URL": "https://doi.org/10.1038/s41598-026-44858-x", "language": "en", "issued": { "date-parts": [ [ 2026, 3, 21 ] ] } }

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Add the badge to the code's README

If you are an author or a maintainer of this paper's code, you can add this badge. It says what OSCR knows of the paper (its code, its validated tracing map) and links the code to this page. Paste one of these snippets under the README's title. Keep the two marks around it, so that the badge can be replaced in place and is never added twice. OSCR changes nothing in your repository.

Markdown

<!-- oscr-badge:start -->
[![OSCR](https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x.svg)](https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x)
<!-- oscr-badge:end -->

reStructuredText

.. oscr-badge:start

.. image:: https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x.svg
   :target: https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x
   :alt: OSCR

.. oscr-badge:end

HTML

<!-- oscr-badge:start -->
<a href="https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x"><img src="https://openscicode.org/badge/doi/10.1038/s41598-026-44858-x.svg" alt="OSCR"></a>
<!-- oscr-badge:end -->

On GitHub, open the README, choose Edit (the pencil), paste the snippet under the title, then "Commit changes".

If you choose to, you can have the badge kept up to date for you, with oscr badge --write from the command line or with the GitHub Action in your repository's workflows. Every other form and option (a card, AsciiDoc, LaTeX, shields.io) is on the badge page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: code errors and mismatches between the code and the paper about this paper and its code, from readers and its authors, beside the paper and its tracing map.

Report a mismatch between the code and the paper · Report a code error (signed in)

Reproductions: reports from readers who ran the authors' code and did not get the paper's result, what they ran, with which environment, commit and data.

Report a reproduction failure (signed in)

Versions of its code: the releases of its code tied to a version of this paper (the preprint, the accepted manuscript, the version of record…), each with the tracing map versioned with it.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.

Memory efficient training for 3D brain image registration networks using PatchMorph.

Code ↔ Paper

Paper

The authors' code

Overview

Abstract

Repositories

cwmok/LapIRN

MungoMeng/Registration-NICE-Net

MungoMeng/Registration-NICE-Trans

BrainImageAnalysis/INRsRegExp

BrainImageAnalysis/PatchMorph

Tracing map

Data

Datasets cited

Code and data availability statement

Versions

Version 1, 30 September 2026: the first record

Cite

This paper

Similar papers

Contribute

Claim this paper

Correct its record

Validate its tracing map

Add the badge to the code's README

Request its removal

Discussion, reproductions, activity