Implementation of a PIF layer in PyTorch accompanying the submission of the manuscript.
Authors: Fabian Eitel, Jan Philipp Albrecht, Martin Weygandt, Friedemann Paul, Kerstin Ritter
Abstract: Neuroimaging data, e.g. obtained from magnetic resonance imaging (MRI), is comparably homogeneous due to (1) the uniform structure of the brain and (2) additional efforts to spatially normalize the data to a standard template using linear and non-linear transformations. Convolutional neural networks (CNNs), in contrast, have been specifically designed for highly heterogeneous data, such as natural images, by sliding convolutional filters over different positions in an image. Here, we suggest a new CNN architecture that combines the idea of hierarchical abstraction in neural networks with a prior on the spatial homogeneity of neuroimaging data. Whereas early layers are trained globally using standard convolutional layers, we introduce for higher, more abstract layers, patch individual filters (PIF). By learning filters in individual latent space patches without sharing weights, PIF layers can learn abstract features faster and regionally specific. We thoroughly evaluated PIF layers for three different tasks and data sets, namely sex classification on UK Biobank data, Alzheimer's disease detection on ADNI data and multiple sclerosis detection on private hospital data. We demonstrate that CNNs using PIF layers result in significantly faster convergence while obtaining the same predictive performance as a standard CNN.
The code for the experiments carried out in the study can be found in the experiments directory. The Jupyter notebook Training and Inference.ipynb can be used to repeat most of the experiments from the study.
Usage:
The pif.py file contains the Patch Individual Filter layer class which can be incorporated as any other layer type.
class MyNet(nn.Module):
def __init__(self):
super(MyNet, self).__init__()
self.conv_1 = nn.Conv3d(1, 8, kernel_size=3, stride=1, padding=0)
self.pool_1 = nn.MaxPool3d(kernel_size=3, stride=1, padding=0)
self.conv_2 = nn.Conv3d(8, 16, kernel_size=3, stride=1, padding=0)
self.pool_2 = nn.MaxPool3d(kernel_size=3, stride=1, padding=0)
self.pif = PatchIndividualFilters3D([15,20,15],
filter_shape=(3,3,3),
patch_shape=(5,5,5),
num_local_filter_in=16,
num_local_filter_out=6,
overlap=1,
reassemble=False,
debug=False)
def forward(self, x):
....The default version of the PIF layer contains only a single Convolution. One can increase the depth of the PIF layer by adding further operations in:
patch_individual_filter_layer/pif.py
Line 78 in 31f5f37