Page 29 - AIH-2-4

Artificial Intelligence in Health Early Parkinson’s detection through CNNs

It can also help develop treatments and identify patients using partial least squares followed by classification into
eligible for therapeutic clinical trials. 5 controls and Parkinsonism by means of a support vector
29
Single photon emission computed tomography machine (SVM) classifier. Illan et al. also used voxels
(SPECT) imaging using 123 I-Ioflupane (DaTSCAN or corresponding to the striatum to train an SVM classifier
[123I]FP-CIT) has been shown to increase the diagnostic with a linear kernel to classify controls and PS. Rojas
30
accuracy of PD, mainly in the earlier stages of the disease, et al. used voxels corresponding to the striatum and then
carried out feature reduction through principal component
by showing the functional deterioration or dopaminergic analysis followed by classification using SVM. Towey
deficit in the striatal region of the brain (which is one of 31
the primary regions affected in PD). The accuracy of et al. performed feature extraction on all voxels through
6-9
diagnosis of PD at an early phase is the poorest based on singular value decomposition followed by classification
32
clinical indices as early symptoms are mild/moderate, into PS or non-PS. Huertas-Fernández et al. calculated
unlike in advanced stages of the disease. Furthermore, the bilateral caudate and putamen uptake and asymmetry
4,5
these symptoms are common in other neurodegenerative indices from SPECT images and developed predictive
disorders like essential tremor and multiple system models using logistic regression, SVM, and LDA to classify
PD from vascular Parkinsonism. Kim et al. used image
20
atrophy, which often leads to misdiagnosis. 10-12 The effects augmentation to increase the size of data and a classifier
of misdiagnosis are severe as it may lead to unnecessary based on the Inception v3 model that can classify normal
medical examinations and therapies, and associated side- from abnormal SPECT scans.
effects. Recent studies have shown that around 3.6 – 19.6%
of clinically diagnosed PD subjects show no dopaminergic There are also many studies using the SPECT data from
deficit, and these subjects are classified as scans the Parkinson’s progression marker initiative (PPMI),
without evidence of dopaminergic deficit (SWEDD). 10- which is among the most popular, widely used, and largest
12 Subsequent follow-up on these subjects have shown database for PD, 11,19,21-27,33,34,45-48 and the same data were used
11
that they neither deteriorate nor respond to levodopa (a in the present study. Choi et al. trained a convolutional
primary medication in PD) and that their SPECT scans neural network (CNN), which they called PD net, using
remain normal in the follow-up imaging. Thus, these SPECT images to classify PD from normal and non-
subjects were considered highly unlikely of having PD and Parkinsonism tremor. They also used the model to classify
that the initial diagnosis of PD was incorrect. 13-15 These SWEDD subjects. In their analysis, they used the complete
studies evidently point out that dopaminergic imaging is volume data, rather than considering a selected range of
highly useful and that an abnormal imaging, at least in slices, due to which the CNN network became complex
cases of diagnostic uncertainty, is strongly supportive of a with many layers. Martínez-Murcia et al. also used a
22
diagnosis of neurodegenerative Parkinsonism (PS), such CNN to differentiate PD from others (healthy normal
as PD. and SWEDD). They used a threshold-based approach to
select sub-volumes from the volume which they later input
In clinical practice, SPECT images are usually analyzed
by visual inspection and/or by region of interest (ROI) to the CNN. They observed that due to this sub-volume
selection, the complexity of the CNN became small with
16
analysis. The visual analysis relies on the judgment of the just two convolutional layers. Martínez-Murcia et al.
21
22
observer that heavily depends on his expertise, experience, used the features extracted from SPECT images through
and knowledge. ROI techniques involve outlining or independent component analysis to train an SVM classifier
17
positioning the ROI over the striatum (target region) and to distinguish PD from normal. They observed much better
the occipital cortex (reference region), and a quantitative performance than their previous work using the voxel-
measure termed the background subtracted striatal uptake as-features approach. Hirschauer et al. used data from
19
ratio is computed. Despite the odds, the latter method different clinical examinations and SPECT imaging, and
6
or the quantitative method is the most acceptable one, trained an enhanced probabilistic neural network model
since, according to a trial study, it provides an excellent to differentiate PD from SWEDD. Oliveira and Castelo-
intra- and inter-observer agreement. However, the ROI- Branco used voxels as features that were extracted based
18
23
based approach relies on manual intervention for placing on volumes of interest defined (which required manual
the ROIs. intervention), and an SVM classifier was used to classify
There have been many studies that make use of PD from normal. The standard binding potential features
machine learning techniques to develop predictive models along with other features related to the volume and length
from SPECT imaging features for the early detection of of the striatal region from SPECT images was utilized to
PD. 11,19-31 Segovia et al. extracted voxels corresponding train an SVM classifier that could classify PD from healthy
28
to the striatum and performed data decomposition normal. Ortiz et al. extracted features from isosurfaces
24
33
Volume 2 Issue 4 (2025) 23 doi: 10.36922/AIH025040005