Page 35 - AIH-2-4

Artificial Intelligence in Health Early Parkinson’s detection through CNNs

Table 3. Performance metrics obtained for different methods for using single slice image and mean image
Method Confusion matrix Accuracy AUC APR Precision Recall Specificity
A. Single slice image
SVM 335 11 95.39 98.70 95.87 96.26 96.82 92.57

13 162
Log Reg 336 10 95.58 98.62 96.25 96.28 97.11 92.57

13 162
MLP 338 8 97.69 99.57 99.11 98.83 97.69 97.71

4 171
CNN 340 6 98.27 99.78 99.45 99.13 98.27 98.29

3 172
B. Mean image
SVM 338 8 96.73 98.78 96.30 97.41 97.69 94.86

9 166

Log Reg 338 8 96.16 98.74 96.95 96.57 97.69 93.14

12 163
MLP 338 8 97.50 99.12 96.56 98.54 97.69 97.14

5 170
CNN 340 6 98.46 99.91 99.80 99.41 98.27 98.86

2 173

Notes: The confusion matrix is represented as True positiveFalse negative . Single slice is the 41 slice image and mean image is the image
st

False positive True negative
obtained after taking the mean of all slices numbered from 35 to 48 in the 3D scan. All values are expressed in percentage (%).
Abbreviations: AUC: Area under the region operating characteristic curve; CNN: Convolutional neural network; LogReg: Logistic regression;
MLP: Multilayer perceptron; SVM: Support vector machine.
Table 4. Classification of the SWEDD data using different 4. Limitations and future work
methods
Recent research shows that deep learning techniques
Method Mean image Single slice image such as the CNN could benefit from the latest advances
True False True False such as data augmentation, which represents a technique
negative positive negative positive used to increase the training data using information
CNN 75 5 76 4 from the available training data. 42 Traditional
LogReg 73 7 73 7 transformations which include a combination of various
LinearSVM 73 7 73 7 affine transformations and using generative adversarial
MLP 74 6 73 7 networks are effective ways to augment the data. Label
43
Abbreviations: CNN: Convolutional neural network; LogReg: Logistic smoothing is another advancement that has shown to
regression; MLP: Multilayer perceptron; SVM: Support vector machine; improve the performance of deep learning models. In
44
SWEDD: Scans without evidence of dopaminergic deficit.
label smoothing, the hard class labels are converted to
CNNs, in this domain, as these models are capable of soft labels. Both data augmentation and label smoothing
learning subtle and complex patterns from training data are methods for regularizing the neural network models,
and making inferences that may even precede clinical which can help in preventing overfitting and also help
judgment. networks in converging faster.

Volume 2 Issue 4 (2025) 29 doi: 10.36922/AIH025040005