Page 30 - AIH-2-4

Artificial Intelligence in Health Early Parkinson’s detection through CNNs

computed from the ROI and trained a CNN-based model relied on full volumes or manual feature engineering, this
to classify PD from healthy normal. Prashanth et al. work uses only the most relevant slice(s) from SPECT
27
computed shape- and surface-fitting-based features and volumes thereby reducing model complexity and the
used machine learning methods to develop classification risk of overfitting. The model was trained and evaluated
models to differentiate scans with deficit, as in PD, on data from the PPMI, one of the most extensive
from scans without deficit, as in normal and SWEDD. and standardized PD imaging databases available. By
Prashanth et al. also used data from multiple modalities incorporating SWEDD into the classification task, this
26
including clinical examinations, laboratory examinations, work contributes toward differential diagnosis within the
and dopaminergic imaging, and developed classification Parkinsonian spectrum. The proposed method combines
models to distinguish early PD from normal. The same the strengths of automated feature learning with informed
researcher group had used the striatal binding ratios to slice selection, enabling improved diagnostic accuracy and
develop classification and prognostic models for PD. practical utility for early-stage PD and SWEDD detection.
25
Zhang et al. employed multimodal data which included
34
SPECT imaging data to identify different PD subtypes 2. Materials and methods
through the long-short term memory (LSTM) networks 2.1. Dataset details
and dynamic time warping. Shiiba et al. extracted
45
radiomics features including intensity- and texture- The data used in the study were obtained from the
based features in the caudate, putamen, and pallidum Parkinson’s Progression Markers Initiative (PPMI)
volumes of interest from the SPECT images and used database (http://www.ppmi-info.org/data). For up-to-date
machine learning methods to classify PD from normal. information, please visit http://www.ppmi-info.org.
Tufail et al. developed a 3D CNN model (consisting of The PPMI is a landmark, large-scale, comprehensive,
46
14 layers including 5 convolution, 5 max pooling, and observational, international, multi-center study that
3 fully connected layers) that is capable of performing recruits de novo (early-untreated) PD patients, and
multiclass classification of Alzheimer’s and PDs using age- and gender-matched healthy subjects to identify PD
positron emission tomography and SPECT neuroimaging progression biomarkers. 4,35
modalities. Majhi et al. used magnetic resonance imaging In this work, SPECT imaging data from the screening
47
and SPECT imaging data to train many deep learning visits of 209 healthy normals, 443 early PD, and, 80
models including VGG16, DenseNet, DenseNet-LSTM, SWEDD were used. All the subjects in the three groups are
and InceptionV3 that are optimized through gray wolf age- and gender-matched to minimize demographic bias.
optimization. Khachnaoui et al. trained deep learning Table 1 shows the age, gender, and Hoehn and Yahr (HY)
48
models based on EfficientNet-B0, Mobilenet-V2, and stage distribution for the three groups. All PD patients were
a custom CNN with 10 layers (4 convolutional, 4 max in their early stage (HY stage 1 or 2 with mean ± standard
pooling) using SPECT images. deviation as 1.50 ± 0.50 ) and all SWEDD subjects exhibited
36
However, several limitations that persist in prior work are early-stage PD symptoms (HY stage as 1.46 ± 0.53).
as follows:
• Use of entire image volumes, increasing model 2.2. Image pre-processing by PPMI
complexity and the risk of overfitting; All the SPECT scans taken at different PPMI sites undergo a
• Dependence on explicit feature extraction pipelines standard pre-processing procedure before they are publicly
followed by machine learning classifiers; shared through the database. This pre-processing was
35
• Manual intervention for ROI placement, reducing carried out so that all scans were in the same anatomical
reproducibility and scalability; alignment (spatially normalized). The process includes
• Training on small subject cohorts, limiting reconstruction from raw projection data and attenuation
generalizability; correction, followed by application of a standard Gaussian
• Focus solely on binary classification (e.g., PD vs. 3D 6.0 mm filter and then normalization of these images to
healthy), with limited attention to diagnostically standard Montreal Neurologic Institute space. These pre-
challenging cases such as SWEDD. processed scans, which were then shared for public access,
The present study addresses the above limitations were used for this analysis. The analysis pipeline is shown
by developing a compact CNN-based model optimized in Figure 1.
through Bayesian hyperparameter optimization to
distinguish early PD from healthy controls, as well as to 2.3. Slice selection
differentiate diagnostically challenging SWEDD cases Each SPECT scan consists of 91 transaxial slices (from
from those with early PD. Unlike prior approaches that bottom to top of the head) each of size 109 × 91, which

Volume 2 Issue 4 (2025) 24 doi: 10.36922/AIH025040005