Alzheimer’s Disease Early Detection Using a Densenet-121
7
It seems that there is a race to obtain greater accuracy, although this metric is misleading. In addition, multiclass classification is avoided. Most studies
implement one-vs-one classifiers, thus achieving higher accuracy values. When
the number of classes increases the accuracy tends to decrease. In fact, we did
not find any article with multiclass classification with more than four classes.
Nor did we find many articles that used the densenet architecture. Only three
papers used densenets, of which two [6,9] are three-dimensional but with shallow densenets and one [11] uses deep densenets but two-dimensional. Finally, the
quantitative analysis of the collected items does not generate a great contribution due to these defects. However, in the review of the articles, we find articles
of remarkable quality as [2]. We also consider that some of the papers collected
are not repeatable.
In contrast to existing studies, we seek to create a multiclass neural network
using only tools available for free. Besides, we do not give greater importance
to accuracy over other metrics and analysis. Finally, we want our process to be
repeatable, and we report it complete along with all the parameters used, as
explained in the next sections.
4 Methodology
In this section, we describe how we collect data using the ADNI study and how we
preprocess these data. Then, we present the development carried out and how we
produced, using the Google Collaboratory tool, an Alzheimer’s prediction model
to fulfill the objective of measuring the accuracy of the detection of Alzheimer’s
disease using a three-dimensional Densenet-121.
4.1 Data Acquisition
In this work, we used the data from ADNI. We used their beta advanced
search functionality with the following criteria. In Projects, we checked ADNI.
In Research Group, we checked MCI, EMCI, AD, SMC, and CN. In Modality,
we checked MRI. We only chose MRI and did not add PET because of economic restrictions. PET requires radiopharmaceuticals, as mentioned. It is more
usual to find MRI in contexts of economic limitations. Continuing with search
options, in Image Description, we used MPRAGE. In Acquisition Plane, we used
SAGITTAL, and finally, in Weighting, we used T1. The rest of the search fields
were left with their default values. With those parameters, we obtained 5556
magnetic resonance images with the following distribution: 1520 Cognitive Normal (CN), 186 Significant Memory Concern (SMC), 1222 Early Mild Cognitive
Impairment (EMCI), 1274 Mild Cognitive Impairment (MCI), 636 Late Mild
Cognitive Impairment, and 718 Alzheimer’s Disease.
The images obtained from ADNI are in Digital Imaging and Communication
On Medicine (DICOM) format. The files are in a zipped archive of 55.5 GB,
and the uncompressed files measure 138 GB. We reduce that size with data
preprocessing, and we explain how and why in the next section.
7
It seems that there is a race to obtain greater accuracy, although this metric is misleading. In addition, multiclass classification is avoided. Most studies
implement one-vs-one classifiers, thus achieving higher accuracy values. When
the number of classes increases the accuracy tends to decrease. In fact, we did
not find any article with multiclass classification with more than four classes.
Nor did we find many articles that used the densenet architecture. Only three
papers used densenets, of which two [6,9] are three-dimensional but with shallow densenets and one [11] uses deep densenets but two-dimensional. Finally, the
quantitative analysis of the collected items does not generate a great contribution due to these defects. However, in the review of the articles, we find articles
of remarkable quality as [2]. We also consider that some of the papers collected
are not repeatable.
In contrast to existing studies, we seek to create a multiclass neural network
using only tools available for free. Besides, we do not give greater importance
to accuracy over other metrics and analysis. Finally, we want our process to be
repeatable, and we report it complete along with all the parameters used, as
explained in the next sections.
4 Methodology
In this section, we describe how we collect data using the ADNI study and how we
preprocess these data. Then, we present the development carried out and how we
produced, using the Google Collaboratory tool, an Alzheimer’s prediction model
to fulfill the objective of measuring the accuracy of the detection of Alzheimer’s
disease using a three-dimensional Densenet-121.
4.1 Data Acquisition
In this work, we used the data from ADNI. We used their beta advanced
search functionality with the following criteria. In Projects, we checked ADNI.
In Research Group, we checked MCI, EMCI, AD, SMC, and CN. In Modality,
we checked MRI. We only chose MRI and did not add PET because of economic restrictions. PET requires radiopharmaceuticals, as mentioned. It is more
usual to find MRI in contexts of economic limitations. Continuing with search
options, in Image Description, we used MPRAGE. In Acquisition Plane, we used
SAGITTAL, and finally, in Weighting, we used T1. The rest of the search fields
were left with their default values. With those parameters, we obtained 5556
magnetic resonance images with the following distribution: 1520 Cognitive Normal (CN), 186 Significant Memory Concern (SMC), 1222 Early Mild Cognitive
Impairment (EMCI), 1274 Mild Cognitive Impairment (MCI), 636 Late Mild
Cognitive Impairment, and 718 Alzheimer’s Disease.
The images obtained from ADNI are in Digital Imaging and Communication
On Medicine (DICOM) format. The files are in a zipped archive of 55.5 GB,
and the uncompressed files measure 138 GB. We reduce that size with data
preprocessing, and we explain how and why in the next section.