Page 33 - AIH-2-4

Artificial Intelligence in Health Early Parkinson’s detection through CNNs

The hyperparameters in a CNN model need to be Table 3 shows the 10-fold cross-validation performance
fine-tuned for optimal performance and to prevent metrics for all methods applied to both single slice and
overfitting. For instance, increasing the number of filters mean image cases. All models demonstrated excellent
in the convolutional layers can help capture more diverse classification performance, with CNN achieving the highest
features, but also leads to higher computational costs and accuracy. The performance measures obtained for the
training time. Moreover, beyond a certain point, additional mean image consistently gave better results as compared
filters may contribute to only minimal improvements in to a single slice, except for the MLP model. This is because
performance. the mean image likely provides a more comprehensive
representation of the striatal region by integrating
2.6.2. Fine-tuning: Hyperparameter optimization information across multiple slices, thereby smoothing out
Model fine-tuning is important for achieving optimal noise and capturing more consistent patterns relevant to
performance. In this study, all models were subjected early PD detection. This richer representation can help
to hyperparameter optimization. For SVM and logistic these models generalize better and improve classification
regression, a grid search combined with cross-validation accuracy.
was employed to identify the best configuration. For the The metrics obtained here significantly improve the
neural network-based models – MLP and CNN – a more results obtained elsewhere and other closely related
27
sophisticated approach was adopted. works. 11,19-31,33,34 For instance, in Prashanth et al.’s work,
27
Research has shown that Bayesian hyperparameter a classification model was developed for the detection of
optimization is more efficient than manual, random, or grid early PD from normal controls using features extracted
search-based methods, particularly for neural networks, from SPECT images, achieving an accuracy of 97.29% and
both in terms of performance and the computation an area under the region operating characteristic curve
time required to identify optimal hyperparameters. (AUC) of 99.26%. In contrast, our approach achieved an
40
In Bayesian optimization, unlike in random search, it accuracy of 99.08% and an AUC of 99.93% using single
keeps track of past evaluation scores which is used to slice images – surpassing previous benchmarks. Notably,
form a probabilistic model mapping hyperparameters to our method does not rely on any explicit feature extraction.
a probability of a score on the objective function p(y|x). Instead, CNN automatically learns discriminative features
Now this probabilistic model is much easier to optimize directly from the input data through its convolution and
than the original objective function, thereby helping in pooling operations, highlighting its capacity for effective
finding the next best set of hyperparameters to evaluate. In end-to-end learning.
this paper, the Tree-structured Parzen Estimator Approach 3.1. Error analysis
was used to estimate the probabilistic model. The optimal
40
architecture for CNN and MLP was estimated based on Among all the attempted methods, CNN gave the best
this optimization and is presented in the Results section. performance metrics. However, as observed in Table 3, very
few records were misclassified. Figure 4 shows examples
3. Results and discussion of images that were misclassified, with Figure 4A showing

The hyperparameters of classification algorithms were a SPECT image from a normal subject but classified as
estimated separately for both cases, which are single slice PD, and Figure 4B showing a SPECT image from a PD
image and mean image, using a hold-out set (Partition 2) subject but classified as normal. It should be noted that
a normal scan is characterized by intense, uniform, and
which was not used in either training or evaluation of the symmetric high-intensity regions on both hemispheres
models. Table 2 shows the estimated hyperparameters.
that appear as two comma-shaped regions (as observed in
The hyperparameters estimated vary between the single Figure 2A). In the case of Figure 4A, it is observed that
slice and mean image cases. The regularization parameter the tail or the bottom of the comma-shaped region is less
C for SVM and logistic regression increased slightly for the intense as compared to the upper region. This might be an
mean image case. This indicates that the models benefited interesting case of misdetection from the CNN model as
from less regularization on the averaged data, and this is the model is actually detecting the non-uniformity in the
because averaging could have led to a reduction in noise comma-shaped region in the image. Training the network
and better overall representation of the pattern needed for with more images like these can help alleviate these errors.
detection. The CNN configurations estimated in the study In fact, such errors may even assist clinicians by flagging
are much more compact and efficient as compared to a potentially ambiguous or borderline cases. Similarly,
related work where five convolutional layers and three Figure 4B is a case of misdetection where an early PD
24
fully connected layers were used. case is detected as normal. Here as well, it is an interesting

Volume 2 Issue 4 (2025) 27 doi: 10.36922/AIH025040005