Table 10 A comparison of medical imaging papers
Article | Main idea | Tools | Applied algorithms | Advantages | Disadvantages |
|---|---|---|---|---|---|
PD diagnosis using diffusion MRI and CNN | •SPSS software | •Greedy algorithm • CNN | • High AUC | • Low scalability • Not utilizing various scanners | |
ML-based model for PD cognitive impairment diagnostics | • FSL • Python | • RF • XGBoost • LightGBM • V-SHARP algorithm • Grid search algorithm | • High accuracy | • Insufficient validation • Low scalability | |
CNN framework for SNPC segmentation | • MATLAB • FSL • SPM • MRtrix3 • R programming • Python (Scikit-learn) | • CNN | • High accuracy • High AUC • Reduced processing time for large datasets • High reproducibility | • No external validation • Not evaluating the effects of medicines on neuromelanin • No SNPC topography analysis • Low scalability | |
PD symptoms impact on HRQoL and neural network mapping | • Python • (Scipy, Python factor_analyzer) • MATLAB • FSL • ART | • Multiple regression • RF | • A thorough analysis of the various factors affecting HRQoL | • Not analyzing parameters such as accuracy and sensitivity • Lack of sleep-related questions • Lack of longitudinal studies about patients | |
DL-based PD diagnosis via MRI and locus coeruleus classification | • Not mentioned | • CNN | • High precision | • Not analyzing parameters such as sensitivity • Considering the limited amount of data | |
DL-based motor performance prediction in PD | • Python | • CNN | • High-performance prediction | • Low scalability • Not evaluating metrics such as sensitivity and specificity | |
Deep complex neural networks for classifying IPD | • MedCalc • R programming | • CNN • YOLOv3 algorithm | • High diagnostic performance • High accuracy | • Not investigating other CNNs • Low scalability | |
Rs-fMRI and topological ML for diagnosing PD | • MATLAB • Python | • UMAP • DNN • SVM • GB | • High accuracy | • Lack of high-quality data • Insufficient analysis of scanning device and embedded parameters • Not evaluating classification performance of different stages of PD | |
DCNN model for dementia diagnosis via FDG-PET | • Python | • DCNN • GAN • Adam optimizer | • High accuracy • High specificity • High sensitivity | • Not investigating other types of dementia • Not assessing non-imaging features | |
Interpretable DL model for PD diagnosis | • ANT | • CNN Jacobians • N3 • SmoothGrad | • High accuracy • High AUC(ROC) • High precision • High sensitivity • High specificity | • Not investigating other DL architectures • Lack of utilization of multimodal MRI data | |
3D and 2D CNN to distinguish PD from healthy | • Python | • 2D CNN • 3D CNN | • High AUC • High accuracy | • Not applying parallelization to increase performance | |
Detection of PD using parameter-weighted matrices | • MATLAB | • CNN | • High AUC (ROC) | • Not able to detect at early PD stages • Lack of varied data • Not verifying the model’s effectiveness externally | |
ML-based PD diagnosis via I-123 FP-CIT scans | • MATLAB | • SVM | • High accuracy • High sensitivity • High specificity | • Applying a relatively small number of patients • Not considering additional parameters • Limited evaluation of diverse methods | |
[18 F] DOPA PET/CT and CNN for PD classification | • Python | • CNN | • High accuracy • High specificity • High sensitivity | • Not performing a regional analysis • No neuropathologic confirmation | |
DL model for PD/HC separation via FDG-PET | • MATLAB | • Radiomatic DL | • High accuracy | • Lack of general data such as race, and nationality • Low scalability • Not involving other kinds of data, such as MRI | |
ML-based PD detection using 123I-ioflupane images | • Not mentioned | • Gradient boosted trees • LR • KNN | • High AUC | • Not considering personal information such as gender • Low scalability | |
DL model for PD detection via facial expression | • Python (PyTorch) | • StarGAN • CNN | • High accuracy | • Limited database diversity | |
DL approach for diagnosing PD from handwriting | • Not mentioned | • ResNet50 • VGG19 • INCEPTION-V3 • KNN | • High accuracy • High precision | • Not analyzing F1-score and specificity | |
Automatic PD subtype diagnosis using SVM | • Python (Scikit-learn) | • SVM • LASSO • SHAP | • High AUC | • Low scalability • No analysis of genetic biomarkers | |
DL-based PD classification using MRI data | • Not mentioned | • ResNeXt | • High accuracy | • Lack of clinical validation | |
DL framework for PD classification via MRI | • Python (PyTorch) | • CNN • Adam optimizer | • High specificity • High accuracy • High sensitivity • High F1-score | • Existence of uncertain details in the model of DL • Low scalability • Not evaluating 3D imaging data | |
PD identification via DL on T1-weighted and QSM scans | • Python (PyTorch) • R programming | • SE-ResNeXt50 • CNN | • High AUC • High accuracy | • Limited number of centers • Low scalability • Reduced performance due to inaccurate segmentation | |
DNN-based PD diagnosis using SPECT images | • Python (Scikit-Learn) • OpenCV | • PARNet | • High specificity • High accuracy • High sensitivity • High F1-score • High precession | • Low scalability • Not examining a large domain of patients | |
PD classification with CNN using DaTSCAN images | • Not mentioned | • MobileNet-V2 • EFFICIENTNET-B0 | • High accuracy | • Low scalability • Potential overfitting issues | |
PD diagnosis with GNNs through MRI scans | • Python (PyTorch) | • GNN • Sparsity ATopk model | • High F1-score | • Not evaluating multimodal data | |
DCNN-based PD diagnosis using TCS images | • ImageJ software • Python (PyTorch) • R programing | • DCNN | • High accuracy • High PPV • High F1-score • High sensitivity | • Low scalability • Retrospective and single-center nature • Not evaluating different PD subtypes | |
The impact of CNN design and data leakage on PD diagnosis using MRI | • 3D Slicer • FSL | • CNN • N4 algorithm | • High accuracy | • Unequal dataset conditions • Limited to T1-weighted MRI • Not using multicenter or multimodal approaches | |
DL-based PD detection using retinal fundus images | • Not mentioned | • LR • SVM • Elastic Net • ResNet50 • Inception-V3 • GoogleNet • VGG-16 | • High NPV • High sensitivity | • Low scalability |
