Page 284 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

6. Conclusions 277

once again for the classification of patients affected by Alzheimer disease. In
Ref. [33], a framework based on deep convolutional neural networks is presented
for infant brain segmentation that combines feature maps deriving from T1, T2
and diffusion-weighted MR images. The authors showed that the integration of multi-
modality images improved the performance compared to previous methods. One
limitation for the applicability of this approach to neuroimaging data is the small
quantity of available samples compared to the number of input features (curse of
dimensionality). However, this problem can be alleviated with a preliminary feature
selection step that in this field is often attained by identifying regions of interest or
by segmenting the brain in parcels and then deriving a single feature from each
parcel.

6. CONCLUSIONS
The need to analyze the growing amount of biomedical multiview data, such as those
made available by high throughput omics technology and brain imaging substantially
increases the importance of machine learning and data integration techniques. The
first attempts to integrate data consisted in merging all views together and performing
a joint data analysis. In the last years, data scientists started developing new method-
ologies that allow to learn from multiple views by taking into account their diversity.
In this work, we reviewed several integration methods that have been successfully
applied to solve biomedical problems. The discussed methodologies were categorized
based on three fundamental aspects of data integration: (1) the type of data under the
analysis (homogeneous or heterogeneous), (2) the statistical problems to solve, and
(3) the stage of the analysis when the integration is performed (early, intermediate,
or late). Examples of how these methods can be applied to solve different research
problems in bioinformatics and neuroinformatics were reported: we discussed the ap-
plications of multiview clustering and classification methodologies to drug reposition-
ing and patient stratification; we reported an example of how both clustering and
classification can be combined in a multiview setting for the automated diagnosis
of neurodegenerative disorders; we discussed how multiple noninvasive imaging mo-
dalities can be exploited together to obtain more accurate brain parcellations. More-
over, we explained how the new emerging deep learning methodologies can be
applied to the biomedical field for multimodal feature learning. As showed in this
work and in many others, significant work has been carried out in the field of multi-
view learning and data integration for biomedical applications. The main problem is
the lack of a general criterion to choose a method among the others. Thus, it is
becoming increasingly necessary to create a framework that allows to perform
different types of integrative analysis on different types of data. Such a tool would
be of paramount importance, especially for those who initially want to approach these
new techniques. In conclusion, even if there are some limitations, the results reached
so far are encouraging. This suggests that the applications of multiview learning tech-
nique to big data analysis in the biomedical field is really promising.