Diagnosis and Classification of Hypothyroidism and Hyperthyroidism Based on Machine Learning Techniques رسالة ماجستير

Abstract

The thyroid gland is one of the most important parts of the body, as any imbalance in it leads to many health problems. Thyroid examinations are one of the necessary periodic examinations that must be carried out to check on health, as any trouble in it causes health complications if it is not treated. Therefore, there is an urgent need for early detection of these diseases as well as accuracy in diagnosis, so Artificial Intelligence (AI) approaches can support medical decisions to detect thyroid disease before the disease situation worsens and affects the patient with other problems. This research uses the Machine Learning (ML) Hybrid Model of Neural Networks (NNs) and Genetic algorithms (GAs) [GAs-MLPNNs], and Deep Learning (DL) Model to classify and predict thyroid diseases depending on features and medical images. The datasets used in this research depend on different resources, we used a global dataset, also local datasets were collected from four cities in Palestine, and a global dataset for ultrasound images of the thyroid. The datasets pass-through data preprocessing and feature extraction process. The final step is the application of different AI methods to recognize the patterns in these datasets. Decision-tree (DT), Naïve Bayes (NB), Support Vector Machine (SVM), k-Nearest Neighbor (KNN), Ensemble Methods, Multi-Layer Perceptron Neural Networks (MLPNNs), GAsvi MLPNNs, and Deep learning Convolution Neural Network (CNN) implemented by VGG-16 Model were used to recognize the pattern on the datasets and improve the results of the classification of thyroid diseases with high accuracy. In the first stage, DT, NB, SVM, KNN, Ensemble, and MLPNNs were applied to the global dataset to classify thyroid diseases, and the accuracy results were as follows: 99.5%, 93.3%, 98.2%, 95.1%, 99.6%, and 95.6% respectively. In the second stage, DT, NB, SVM, KNN, Ensemble, and MLPNNs have been applied to the local dataset in two ways, the whole data collected from the four cities, and the data for each city. All cities datasets obtained the highest accuracy when applying the Ensemble model which produces 91.10%. The best model in terms of accuracy of results for each city is DT, which obtained the accuracy; 92.40%, 88.50%, 91.90%, and 89.90% in Ramallah, Nablus, Qalqiliah, and Salfeit respectively. In the third stage, a custom global dataset was created to compare the results of classification to the local datasets. In the fourth stage, Hybrid Model (GAs-MLPNNs) was used to improve the accuracy of the classification of thyroid diseases, the model was applied on a custom global dataset and local dataset, the results were as follows: 95% for the custom global dataset, and 96% for the local dataset. In the fifth and final stage, a VGG-16 model based on CNN was used to classify a group of ultrasound images of the thyroid gland into a malignant tumor or a benign gland tumor, and the accuracy result of the classification was 87.00%.