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
☆
114 Computational Intelligence Algorithms
Speech language therapy improves interaction and communication among other
skills.
Farooq et al. (2023) utilized support vector machine and logistic regression to classify the ASD factors and detection of ASD. Four different datasets were ana­lyzed for children, for which the proposed model achieves 98% accuracy, and for adults, for which 81% accuracy as achieved [1].
Al-disbat et al. (2018) experimented with fuzzy data mining algorithms for ASD kids for classication in the University of California, Irvine Machine Learning (UCIML) repository dataset. The researchers evaluated various models of FUZZY and found the FURIA model produced better results than others [2]. Thabath et al. (2020) proposed rules in the machine leaning (ML) model for classication and showed the actual cause for this disease. The researchers used this model in experiments with children, adolescents, and adults datasets with boosting, bagging, decision trees, and rule induction algorithms and produced higher accuracy [3]. Wingeld et al. (2020) developed an ML embedded mobile application to monitor autism detection and identied that the random forest algorithm AUCROC produced accuracy of 98%, and this application helps in many ways to identify ASD [4]. Leroy et al. (2024) analyzed the ML model, three deep learning (DL) models, and various ensemble models for ASD detection and found that majority voting with BiLSTM Ml achieved 100% precision, 91% accuracy, 100% specicity, and a 0.91 F1 score [5]. Reddy et al. (2023) used a facial image dataset for CNN, VGG16, VGG19, and the EfcientNet BO classication and attained accuracy of 84.66%, 98.50%, and
87.9% for each algorithm [6].
Alsaade et al. (2022) developed a model with web application and DL algorithms such as CNN, Xception, VGG19, and NASNETMobile to classify the facila autism image dataset and attained accuracy of 78% for NASNETMobile, 80% accuracy for VGG19, and 91% accuracy for the Xception model [7]. Khosla et al. (2021) proposed a DL model to classify facial images of healthy and autistic children and removed the duplicate images. The Mobile net model attained the accuracy of 87% in clas­sication [8].
Research gaps include the following:
• Identication and treatment of ASD takes more time, which leads to proper medical practice.
• Many researchers experiment with DL and ML for ASD analysis, yet they struggle to attain high accuracy.
• Many researchers experiment with feature datasets instead of image datasets.
9.6 METHODOLOGY
The failure of ML technique was the basic reason to develop a DL algorithm. DL is an advanced technology of ML and articial intelligence (AI). It can handle large sets of data and doesn’t need external feature extraction techniques. It contains mul­tiple layers of neural networks and hidden layers, and every process is carried out promptly. This architecture can process large data in accurate manner within short
115 Intelligent Deep Learning Algorithms
time. DL algorithms play a signicant role in the medical eld due to its efcient and timely analysis [9]. Every algorithm performs its role in unique way due to its interconnected neurons and ability to learn from the given data [10]. DL models have an ability to capture dominant features and works in an end-to-end fashion. It plays a dominant role in classication, object detection, segmentation, speech recognition, sentiment analysis, medical analysis, predictive analysis, fraud detection, recom­mender systems, etc. [11, 12]. This research analyzed DL techniques such as CNN, RNN-CNN, LSTM-CNN, VGG-16, and ResNet-50 for autism detection.
9.6.1 DATASETS
The facial dataset for ASD is picked from the Kaggle dataset, and it contains 1,468 autism les and 1,468 nonautism les. The training folder contains 1,628 for autism and 1,628 for nonautism. At the same time, the test folder contains 150 for autism and 150 for nonautism images. The validation folder contains 50 autism and 50 nonautism images. A total of 5,874 images were experimented with for this study.
Figure 9.2 shows the autism and nonautism images in the dataset. Image resizing is
carried out for image preprocessing.
Figure 9.3 shows the overall methodology of the ASD classication.
9.6.2 CNN
This network was developed to execute grid information and mainly used for image analysis. Convolution layer, activation function, stride, padding, pooling layer, batch normalization, attening, and fully connected layers are the basic operations and building layers for CNN. Convolution layers works with input and lters by dot prod­uct and produce feature maps by capturing dominant features. Pooling layers are utilized to reduce the dimensionality of feature maps and can be either max pooling or average pooling. Figure 9.4 shows the layers and architecture of CNN [13].
FIGURE 9.2 Autism and nonautism images.
116 Computational Intelligence Algorithms
FIGURE 9.3 Schematic model of workow.
The max pooling fetches the maximum value from the feature map by ltering, and average pooling considers the average value from the feature map region. This research utilized the concept of max pooling. The activation function helps to iden­tify the nonlinearity in the model and learned complex features. Every layer used the concept of the same padding to protect the edge features by adding zero. Fully
FIGURE 9.4 CNN layer architecture.
117 Intelligent Deep Learning Algorithms
Ht = f (H
t−1
, xt)
t1−
= tanh WH
t
)
H
t
(
hh t −1
+ Wx
xh
yt = W
hy
,H
t
y
t
connected layers work as same as dense layers in ML and fetch the nal feature maps with dominant patterns. The activation function sigmoid produces binary output for classication.
9.6.3 RNN
This network processes sequential and temporal data (see Figure 9.5). Hidden layers in RNN capture data from a previous sequence with the help of hid­den layers. It works similar to CNN, and it also has memory to capture pre­vious data. The output depends on previous data. The important property of RNN is the hidden state or memory state, which reduces the complexity of parameters.
The recurrent unit is the basic unit, and it keeps the hidden state. The hidden state holds the knowledge of previous time step and is updated as per the following for­mula. The current state can be calculated by ent state,
is the previous state, and t is the initial setup. The activation function
performs as follows:
where hh is recurrent neuron weight and xh is the input neuron weight. The out­put layer is computed as
, where
is output and hy is the output layer
weight.
Back propagation helps to update these parameters. This process is repeated until the calculation is made for output. The nal output is compared with actual out­put, and bias is propagated back to the network for changing the weight. Hence it remembers every detail of the process. RNN combined with CNN performs image classication based on image features [14, 15].
., where t is the pres-
FIGURE 9.5 RNN
118 Computational Intelligence Algorithms
9.6.4 LSTM
LSTM utilized the concept of RNN, and it has a structure to maintain read, write, and forget states. It reads and writes signicant information and forgets unnecessary information. Developed by Hochreiter and Schmidhuber, the process overcame the problem of vanishing gradient from RNN. It maintains long-term dependencies and consists of three gates: the forget gate, the input gate, and the output gate. The forget gate simply forgets the unwanted information, the input gate captures new informa­tion, and the output gate carries and passes the present time stamp information to the next state. This full process is called single time stamp. Figure 9.6 presents the single time stamp of LSTM.
LSTM networks works on sequential data so that image features were extracted
rst and combined a rchitecture of LSTM with CNN attained t he desired classication.
9.6.5 VGG-16
This DL model is based on the CNN architecture and has 16 convolution layers and is made up of the basic constructive layers of CNN. A. Zisserman and K. Simonyan are the developers of VGG-16.
VGG-16 has one input layer, 13 convolution layers, ve pooling layers, three fully
connected layers, and one output layer.
9.6.6 RESNET-50
ResNet-50 is a combination of CNN architecture with residual blocks. The residual blocks help to recover the degradation problem. ResNet-50 allows the direct delivery of data through skip connections. It acts next to the convolution and batch normal­ization layer. ResNet-50 captures positive values, thereby learning critical patterns of data. The bottleneck convolution layer is special structure of ResNet and is a collection of three convolution layers, the batch normalization layer, and the Relu activation function. These convolution layers use 1 × 1, 3 × 3, and 1 × 1 lters, prevent information loss, and extract dominant features from the data, and the nal lter helps to restore the information. Skip Connections allows unchanged input to the convolution layer output. All information is securely transferred to the next layer without any loss. Hence it performs the deeper learning of networks.
FIGURE 9.6 LSTM gates.
119 Intelligent Deep Learning Algorithms
TruePositive + TrueNegative
A
TruePositiveTrueNegativeFalsePositiveFalseNegative+ + +
True Positive
True Positive + FalsePositive
True Positive
True Positive + FalseNegative
Precision Recall
Precision + Recall
9.7 RESULT AND DISCUSSION
This section summarizes the performance of experimented DL algorithms with proper evaluation metrics. The evaluation metrics accuracy, recall, F1 score, and precision are utilized to discuss the outcome of the framework.
Accuracy: It assess the true prediction from the overall given prediction.
ccuracy = (9.1)
Precision: It shows the correctly predicted positive instances from overall
positive classes.
recision =
(9.2)
Recall: It measures the correct prediction of actual positive classes.
ecall =
(9.3)
F1 score: This determines the harmonic mean between precision and recall
and also shows distribution of unequal classes.
1score = 2
*
*
(9.4)
Table 9.2 and Figure 9.7 show the performance of deep experimented models in
percentage.
Figure 9.8–9.12 demonstrate the accuracy and loss for each classication model.
Using information from Figures 9.8 to 9.12, Figure 9.13 shows the DL model clas­sication performance.
FIGURE 9.7 Performance of deep experimented models.
120 Computational Intelligence Algorithms
TABLE 9.2 Performance of Deep Experimented Models
Models Precision Recall F1 Score Accuracy Loss
CNN 84 89 93 95 1.63 RNN-CNN 79 85 87 90 1.67 LSTM-CNN 80 86 89 92 1.59 VGG-16 81 87 90 94 1.54
ResNet-50 87 92 97 99 1.47
FIGURE 9.8 CNN classication.
FIGURE 9.9 RNN-CNN classication.
FIGURE 9.10 LSTM-CNN classication.
FIGURE 9.11 VGG-16 classication.
121 Intelligent Deep Learning Algorithms
FIGURE 9.12 ResNet-50 classication.
FIGURE 9.13 Accuracy and loss for DL models
122 Computational Intelligence Algorithms
9.8 CONCLUSION
ASD is a silent disease that gets worse with the long-run growth of aficted children. It totally affects the social behavior, understanding, communication, and physical activi­ties of the children. Early identication may control the severeness of the ASD [16–18]. This chapter tries to identify the better classication algorithm for ASD facial image recognition for autistic and nonautistic children. It proposes ve DL models for autism detection, which include CNN, RNN-CNN, LSTM-CNN, VGG-16, and ResNet-50. The CNN model achieves an accuracy of 95%, RNN-CNN attains an accuracy of 90%, 92% accuracy for LSTM-CNN, 94% for VGG-16, and 99% for ResNet-50. As a result ResNet-50 attains the highest result for ASD classication. This chapter nds that every model performs well in classication, and differences between them are very minor, but ResNet outperformed the other models. Hence this chapter provides a solution to identify autistic children at low-cost implementation.
9.9 FUTURE ENHANCEMENT
For ASD detection, a golden metric dataset is not available for public usage. Upcoming researchers need a proper dataset for ASD detection. It will create a gate­way to analyze various advanced architectures with AI integration.
REFERENCES
1. Farooq, M. S., Tehseen, R., & Sabir, M., et al. (2023). Detection of autism spectrum disorder (ASD) in children and adults using machine learning. Scientic Reports, 13(1), 9605.
2. Al-diabat, M. (2018). Fuzzy data mining for autism classication of children. International Journal of Advanced Computer Science and Applications, 9(7), 11–17
3. Thabtah, F., & Peebles, D. (2020). A new machine learning model based on induction of rules for autism detection. Health Informatics Journal, 26(1), 264–286.
4. Wingeld, B., Miller, S., Yogarajah, P., Kerr, D., Gardiner, B., Seneviratne, S., Samarasinghe, P., & Coleman, S. (2020). A predictive model for pediatric autism screening. Health Informatics Journal, 26(4), 2538–2553.
5. Leroy, G., Andrews, J. G., KeAlohi-Preece, M., Jaswani, A., Song, H., Galindo, M. K., & Rice, S. A. (2024). Transparent deep learning to identify autism spectrum disor­ders (ASD) in EHR using clinical notes. Journal of the American Medical Informatics Association, 31(6), 1313–1321.
6. Reddy, P., & J, A. (2023). Diagnosis of autism in children using deep learning techniques by analyzing facial features. Engineering Proceedings, 59, 198.
7. Saleh AY, Chern LH. Autism Spectrum Disorder Classication Using Deep Learning. International Journal of Online & Biomedical Engineering. 2021 Aug 1;17(8).
8. Khosla, Y., Ramachandra, P., & Chaitra, N. (2021). Detection of autistic individu­als using facial images and deep learning. 2021 IEEE International Conference on Computation System and Information Technology for Sustainable Solutions (CSITSS), 1–5.
9. Thabtah, F., Kamalov, F., & Rajab, K. (2018). A new computational intelligence approach to detect autistic features for autism screening. International Journal of Medical Informatics, 117, 112–124.
123 Intelligent Deep Learning Algorithms
10. Shrivastava, T., Singh, V., & Agrawal, A. (2024). Autism spectrum disorder detection with kNN imputer and machine learning classiers via questionnaire mode of screen­ing. Health Information Science and Systems, 12(1), 18.
11. Zhao, F., Ye, S., Zhang, M., Lv, K., Qiao, X., Li, Y., Mao, N., Ren, Y., & Zhang, M. (2023). Multi-classier fusion based on belief-value for the diagnosis of autism spec­trum disorder. Frontiers in Human Neuroscience, 17, 1257987.
12. Thabtah, F., Spencer, R., Abdelhamid, N., Kamalov, F., Wentzel, C., Ye, Y., & Dayara, T. (2022). Autism screening: An unsupervised machine learning approach. Health Information Science and Systems, 10(1), 26.
13. Thamilarasi, V., & Roselin, R. (2020). Automatic classication and accuracy by deep learning using CNN methods in lung chest X-ray image. IOP Conference Series: Materials Science and Engineering, 1055, 012099.
14. Asaithambi, A., & Thamilarasi, V. (2023). Classication of lung chest X-ray images using deep learning with efcient optimizers. 2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC), 0465–0469.
15. Qureshi, M. S., Qureshi, M. B., Asghar, J., Alam, F., & Aljarbouh, A. (2023). Prediction and analysis of autism spectrum disorder using machine learning techniques. Journal of Healthcare Engineering, 2023, 4853800.
16. Alkahtani, H., Aldhyani, T. H. H., & Alzahrani, M. Y. (2023). Deep learning algo­rithms to identify autism spectrum disorder in children-based facial landmarks. Applied Sciences, 13(8), 4855.
17. Simeoli, R., Rega, A., & Cerasuolo, M., et al. (2024). Using machine learning for motion analysis to early detect autism spectrum disorder: A systematic review. Review Journal of Autism and Developmental Disorders. https://doi.org/10.1007/s40489-024-00435-4
18. Erkan, U., & Thanh, D. (2019). Autism spectrum disorder detection with machine learning methods. Current Drug Therapy, 15, 297–308.