Pooling Layer: Each convolutional layer in 3D CNN may contain a pooling layer. Pooling layer simply
takes multiple voxels and produces a single output to the input of the next layer by taking the average
or maximum of the group of input voxels.
In backward pass, the CNN adjusts its weights and parameters according to the output by
calculating the error by means of some loss functions, (other names are cost function and error
function) and backpropagating the error with some rules towards the input. The loss is calculated by
taking the partial derivative of w.r.t. the output of each neuron in that layer such as
for
the output,
of
unit in layer . Chain rule allow us to write the add up the contribution
of each variable as follows:
Weights in the previous convolutional layer can be updated by backpropagating the error to the
previous layer according to the following equation:
Eq. (6) allows us to calculate the error for the previous layer. Also, above eq. makes sense for the only
those points which are n times away from each side of the input data. This situation can be avoided
by simply adding the zero padding to the end of each side of the input volume.
3. 3d medical imaging pre-processing
The preprocessing of the image dataset before feeding the CNN or other classifiers is important
for all types of imaging modalities. Several preprocessing steps are recommended for the medical
images before they are fed as input to the deep neural network model, such as 1) artifact removal, 2)
normalization, 3) slice timing correction (STC), 4) image registration, and 5) bias field correction.
Although all the steps through 1) to 5) help in getting reliable results, STC and image registration are
very important in the case of 3D medical images (especially MR and CT images). Artifact removal
and normalization are the most performed preprocessing steps across modalities. We briefly discuss
the above pre-processing steps.
The first part of any preprocessing pipeline is the removal of artifacts. For example, we may be
interested in removing skulls in brain CT scans before feeding to 3D CNN. Removal of extra-cerebral
tissues is highly recommended before analyzing the T1 or T2 weighted MRI, and DTI modalities for
brain images. fMRI data often contains transient spikes artifacts or is slowed over drift time. Thus,
the principal component analysis technique can be used to look at these spike related artifacts
[3,24,25]. Before feeding the data for preprocessing to an automated pipeline, a manual check is also
advisable. For example, if the input T1 anatomical data is large in size, FSL BET command will not