Page 282 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

5. Deep Multimodal Feature Learning 275

however, they are characterized by high variable phenotypes thus making their
diagnosis challenging. Since a definitive diagnostic test does not exist yet, a model
abletodiscernthedifferentpathologicalmechanismunderlyingthe neurophysiological
signs by identifying reliable diagnostic and prognostic biomarkers would represent
a substantial advance. The results of this work, showing classification accuracies
significantly above chance, suggest that approaches based on integration of multimodal
brain images have the potential of bringing new insights in the investigation of neuro-
degenerative disease.

4.2 MULTIMODAL BRAIN PARCELLATION
To better understand how the human brain works, it is often necessary to subdivide
the brain into parcels in order to understand its modular and hierarchical organization
[26]. For decades now neuroscientists have been seeking for a parcellation of the
human brain consisting of spatially contiguous and nonoverlapping regions that
showed homogeneous characteristic from both the anatomical and the functional
points of view. The main limitation of existing anatomical atlas is that they might
not adapt well to the signal of individual acquisitions, due to intersubject variability.
On the other hand, parcellations that are based on individual brain activity can be
difficult to compare or reproduce on new samples. A great step forward in this context
has been possible thanks to the introduction of a multiview late integration method-
ology proposed by Glasser et al. in 2016 [27]. In this work, multiple imaging
modalities are used to derive several brain maps describing different properties:
cortical myelin content, cortical thickness, pattern of activation from task fMRI, func-
tional connectivity from resting state fMRI. Considering sharp transitions in two or
more of the above measures, a map of potential parcel borders is determined, and
each parcel is described by a multimodal fingerprint, consisting of a set of features
derived from the multiple views. Multimodal fingerprints are then used to train a
classifier to identify parcels in individual subjects. In this way, the authors identified
180 regions per hemisphere. The great innovation introduced by this approach is the
possibility to apply the parcellation to new subjects using the trained model, the only
requirement being the availability of the same set of features for the new samples.

5. DEEP MULTIMODAL FEATURE LEARNING

Classical machine-learning methodologies strongly rely on data representation
based on feature sets. The identification of the right features for a specific task
can be challenging and is usually performed by experts. A new class of deep
machineelearning techniques is emerging that is able to cope with this problem.
Deep learning models are composed of multiple processing layers able to represent
data with a high level of abstraction [28]: each layer learns a representation of input
data based on the output of the previous layer building a hierarchy of ever more
abstract features. The ability of deep neural networks (DNNs) to learn a compressed