Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:

Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_5234_Библиотеки_им_академика_М_И_Перельмана

.pdf
Скачиваний:
0
Добавлен:
15.09.2026
Размер:
16 Мб
Скачать
☆
274 Computational Intelligence Algorithms
convolutional layer, and they are stacked to increase the network’s depth while catching intricate characteristics. The architecture is symmetrical, with many convolutional layers followed by spatial dimension-reducing max-pooling lay­ers. Predictions are made using the nal three layers, which are fully integrated. With around 138 million parameters, VGG16 is a deep model with a lot of param­eters. The VGG19 structure, which has 19 layers with learnable weights, is an expansion of the VGG16 architecture. It has three fully connected layers and 16 convolutional layers; in comparison to VGG16, it has more convolutional layers, which enables more intricate feature extraction. Similar to VGG16, VGG19 makes use of multiple max-pooling layers and 3 × 3 convolutional lters. Classication uses the completely connected layers at the end. Because VGG19 has more lay­ers than VGG16, it has somewhat more parameters, that is, about 144 million. DenseNet121 is comprised of 121 layers, which are arranged into dense blocks where each layer gets input from all layers that came before it in that chunk. The pattern of dense connectedness encourages feature reuse, which leads to a large reduction of parameters. Dense blocks make up the architecture, which is followed by transition layers that manage the network’s complexity and shrink its spatial dimensions. These layers include pooling operations. Compared to conventional deep networks, DenseNet121 is signicantly more efcient because it contains around eight million parameters. DenseNet169 extends the depth to 169 layers while adhering to the same architectural concepts as DenseNet121. Additionally, it makes use of dense blocks, which connect every layer to every layer before it, improving feature propagation and lowering the number of parameters. To control feature map expansion and avoid overtting, the network has multiple transition layers. With almost 14 million parameters, DenseNet169 strikes a fair compromise between model complexity and parameter efciency. The modular architecture of InceptionNetV3 includes “Inception modules” that employ parallel convolutional layers with various lter sizes (e.g., 1 × 1, 3 × 3, 5 × 5) inside the same module. This makes it possible for the model to efciently capture multiscale characteris­tics. To lower computational costs, the network additionally uses factorized convo­lutions, which divide bigger convolutions into smaller ones (e.g., 7 × 7 into 1 × 7 and 7 × 1). With over 23 million parameters, InceptionNetV3 strikes a balance between efciency and depth throughout its several layers.
The foundation of Xception is the concept of depthwise separable convolu­tions. These convolutions drastically reduced the number of parameters by split­ting the typical convolution operation into two parts: a depthwise convolution (spatial ltering) and a pointwise convolution (channel combination). The archi­tecture is composed of several depthwise separable convolutional layers, with fully connected layers for classication at the end. With over 23 million param­eters, Xception is a very efcient model with an architecture that is both simple and robust.
18.4 PROPOSED RESEARCH METHODOLOGY
This section describes the systematic procedure that was employed in this work to use pretrained DL models to identify PD using drawing images. The preparation of
275 Parkinson’s Disease Detection
TP + TN
A
TP + TN + FP + FN
TP
TP + FP
TP
TP + FN
Precision × Recall Precision + Recall
the input, data preprocessing, data splitting, model training, and assessment are the ve primary processes in the approach.
The rst step involves collecting drawing images from the dataset, which includes drawings from both healthy individuals and patients with PD. The dataset contains a total of 204 images, evenly divided between the two classes. Secondly, the raw spiral images were preprocessed, such as resizing them to 224 × 224 pixels to meet pretrained model requirements, converting them to red, green, and blue (RGB) for consistent input data format, and shufing them ran­domly to prevent training biases and ensure good model generalization. Then these preprocessed data were split into training (80%) and testing (20%) sets. The pretrained models are trained using the training set, which consists of 163 images split evenly between the classes for PD and those for healthy people. Each of the six pretrained DL models (VGG16, VGG19, DenseNet121, DenseNet169, InceptionNetV3, and Xception) was trained separately on the preprocessed draw­ing images. The training data were used to ne-tune the models so that they pre­cisely target the goal of differentiating between people who have PD and others who are healthy. The performance of every trained model is evaluated using the testing data consisting of 41 images. A range of performance criteria, including the F1score, accuracy, precision, recall, and ROC curve, are used to assess the models’ effectiveness.
ccuracy = (18.1)
•
recision = (18. 2)
•
ecall = (18.3)
•
•
1-Score2 (18.4) =×
The research methodology for detecting PD using pretrained DL models is illus­trated in Figure 18.3.
FIGURE 18.3 Proposed research methodology.
276 Computational Intelligence Algorithms
18.5 EXPERIMENT EVALUATION AND RESULTS
The experimental work used Jupyter Notebook within Google Colab, a cloud-based platform providing interactive computing resources. Depending on availability, Google Colab offers a free tier with graphics processing unit (GPU) support, speci­cally utilizing NVIDIA Tesla K80, T4, or P100 GPUs. In addition, the hyperparameter tuning involves using the “sigmoid” activation function and “binary_crossentropy” loss function, optimized with “adam.” For 50 epochs, the model is trained with a batch size of 32.
During training, the VGG16 model showed remarkable performance, attaining awless metrics with 100.00% precision, accuracy, recall, and F1 score. It continued to perform well on the testing set, achieving an F1 score of 97.44%, accuracy of
97.56%, precision of 95%, and recall of 100.00%. These outcomes highlight how well VGG16 can categorize photos of drawings to diagnose PD. Figure 18.4 depicts the confusion matrix, Figure 18.5 shows the training and testing accuracy curves, and Figure 18.6 shows the ROC curve.
With an accuracy of 95.09%, precision of 91.21%, recall of 100.00%, and F1 score of 95.40%, the VGG19 model demonstrated strong performance. Testing results for the model showed 90.24% accuracy, 82.61% precision, 100.00% recall, and 90.48% F1 score. Figure 18.7 presents the confusion matrix for VGG19. Fig ure 18.8 shows the training and testing accuracy curves, and Figure 18.9 shows the ROC curve.
FIGURE 18.4 Confusion matrix (VGG16).
FIGURE 18.5 Training and test accuracy curves (VGG16).
277 Parkinson’s Disease Detection
FIGURE 18.6 ROC curve (VGG16).
278 Computational Intelligence Algorithms
FIGURE 18.7 Confusion matrix (VGG19).
FIGURE 18.8 Training and test accuracy curves (VGG19).
FIGURE 18.9 ROC curve (VGG19).
279 Parkinson’s Disease Detection
Training metrics of 99.39% accuracy, 100.00% precision, 98.80% recall, and
99.39% F1 score were attained by the DenseNet121 model. On the testing set, the model achieved 87.80% accuracy, 85.00% precision, 89.47% recall, and 87.18% F1 score. Figure 18.10 displays the confusion matrix, while Figu re 18.11 shows the train­ing and testing accuracy curves., and the ROC curve is presented in Fig ure 18.12.
The training metrics of the DenseNet169 model were 100.00% for accuracy,
100.00% for precision, 100.00% for recall, and 100.00% for F1 score. The model performed as follows on the testing set: 90.24% accuracy, 94.12% precision, 84.21% recall, and 88.89% F1 score. Figures 18.13 and 18.14 depict the training and testing accuracy curves, Figu re 18.15 shows the ROC curve, and Figure 18.13 shows the confusion matrix.
Training metrics of 100.00% for accuracy, 100.00% for precision, 100.00% for recall, and 100.00% for F1 score were attained by the InceptionNetV3 model. It obtained 92.68% accuracy, 94.44% precision, 89.47% recall, and 91.89% F1 score on the testing set. Figure 18.16 displays the confusion matrix, Figure 18.17 shows the training and testing accuracy curves, and Figure 18.18 shows the ROC curve.
With training parameters of 100.00% accuracy, 100.00% precision, 100.00% recall, and 100.00% F1 score, the Xception model performed admirably. On the testing set, it produced an accuracy of 90.24%, precision of 85.71%, recall of 94.74%, and F1 score of 90.00%. The confusion matrix is shown in Figure 18.19, the training and testing accuracy curves are illustrated in Fig ure 18.20, and the ROC curve is presented in Fig ure 18.21.
280 Computational Intelligence Algorithms
FIGURE 18.10 Confusion matrix (DenseNet121).
FIGURE 18.11 Training and test accuracy curves (DenseNet121).
FIGURE 18.12 ROC curve (DenseNet121).
281 Parkinson’s Disease Detection
FIGURE 18.13 Confusion matrix (DenseNet169).
282 Computational Intelligence Algorithms
FIGURE 18.14 Training and test accuracy curves (DenseNet169).
FIGURE 18.15 ROC curve (DenseNet169).
283 Parkinson’s Disease Detection
FIGURE 18.16 Confusion matrix (InceptionNetV3).
FIGURE 18.17 Training and test accuracy curve (InceptionNetV3).