6
B. Solano-Rojas et al.
Positron Emission Tomography. PET scans use radiopharmaceuticals to
create three-dimensional images. These types of scans produce small particles
called positrons. A positron is a particle with roughly the same mass as an
electron but oppositely charged. Positrons react with electrons in the body, and
when these two particles combine, they annihilate each other. This annihilation
produces a small amount of energy in the form of two photons that shoot off in
opposite directions. The detectors in the PET scanner measure these photons
and use this information to create images of internal organs [16].
3 Previous Work
Our literature review assesses how much progress has been made and what can
be contributed in the detection of AD using deep learning, in particular with
Convolutional Neural Networks (CNN). We only focus on AD however detection
of another neurodegenerative disease using DNNs has been investigated [13,19].
We used IEEE
2 as the source for Artificial Neural Networks because, according to Journal Rankings
3 on the category of Artificial Intelligence, IEEE is the
first on both SJR and H-Index sortings. We used the search engines Duck Duck
Go, and Google Scholar to find illustrative publications.
We used the search string “deep AND learning AND alzheimer AND mri”
in order to assess the use of convolutional deep learning in our application of
interest. We ran the query mentioned from 2016 to the present (in 2019) since we
are searching about recent advancements in neural networks. We retrieved from
IEEE Digital Library 81 records with this query, including conferences, journals,
and early access articles.
We screened by title, and if the title was too ambiguous by abstract. We
searched for the application of convolutional deep learning and we obtained 32
articles. Notably, we searched for literature that included the design of convolutional deep learning artifacts for computer vision to detect AD in MRI and
other modalities. Besides, the literature was restricted to supervised learning.
For example, we did not include convolutional autoencoders alone.
For the articles we deemed appropriate, we developed a data extraction
spreadsheet to serve for analysis where we collected the following information
about each publication: (1) year of the paper, (2) architecture of the neural
network, (3) if the MRI images were processed, (5) the modalities (number of
inputs), (6) the number of classes used, and the metrics of (7) accuracy, (8)
sensitivity, (9) specificity, and finally (9) the Area Under the curve Receiver
Operating Characteristics (AUROC).
In this literature review, with our data extraction spreadsheet, we find a
severe problem. Almost 50% of papers report accuracy but do not report sensitivity, specificity or AUROC. Accuracy alone can be misleading. A classifier
can report a high accuracy and yet have a low capacity of true prediction. We
also conclude that the studies are too diverse to allow a meaningful comparison.
2 https://ieeexplore.ieee.org/.
3 https://www.scimagojr.com/.
B. Solano-Rojas et al.
Positron Emission Tomography. PET scans use radiopharmaceuticals to
create three-dimensional images. These types of scans produce small particles
called positrons. A positron is a particle with roughly the same mass as an
electron but oppositely charged. Positrons react with electrons in the body, and
when these two particles combine, they annihilate each other. This annihilation
produces a small amount of energy in the form of two photons that shoot off in
opposite directions. The detectors in the PET scanner measure these photons
and use this information to create images of internal organs [16].
3 Previous Work
Our literature review assesses how much progress has been made and what can
be contributed in the detection of AD using deep learning, in particular with
Convolutional Neural Networks (CNN). We only focus on AD however detection
of another neurodegenerative disease using DNNs has been investigated [13,19].
We used IEEE
2 as the source for Artificial Neural Networks because, according to Journal Rankings
3 on the category of Artificial Intelligence, IEEE is the
first on both SJR and H-Index sortings. We used the search engines Duck Duck
Go, and Google Scholar to find illustrative publications.
We used the search string “deep AND learning AND alzheimer AND mri”
in order to assess the use of convolutional deep learning in our application of
interest. We ran the query mentioned from 2016 to the present (in 2019) since we
are searching about recent advancements in neural networks. We retrieved from
IEEE Digital Library 81 records with this query, including conferences, journals,
and early access articles.
We screened by title, and if the title was too ambiguous by abstract. We
searched for the application of convolutional deep learning and we obtained 32
articles. Notably, we searched for literature that included the design of convolutional deep learning artifacts for computer vision to detect AD in MRI and
other modalities. Besides, the literature was restricted to supervised learning.
For example, we did not include convolutional autoencoders alone.
For the articles we deemed appropriate, we developed a data extraction
spreadsheet to serve for analysis where we collected the following information
about each publication: (1) year of the paper, (2) architecture of the neural
network, (3) if the MRI images were processed, (5) the modalities (number of
inputs), (6) the number of classes used, and the metrics of (7) accuracy, (8)
sensitivity, (9) specificity, and finally (9) the Area Under the curve Receiver
Operating Characteristics (AUROC).
In this literature review, with our data extraction spreadsheet, we find a
severe problem. Almost 50% of papers report accuracy but do not report sensitivity, specificity or AUROC. Accuracy alone can be misleading. A classifier
can report a high accuracy and yet have a low capacity of true prediction. We
also conclude that the studies are too diverse to allow a meaningful comparison.
2 https://ieeexplore.ieee.org/.
3 https://www.scimagojr.com/.