Integration of machine learning and deep learning methods for the enhancement of rheumatoid arthritis diagnosis

In this study to classify normal and Rheumatoid arthritis hand radiographs, convolutional neural networks were used as object detectors and feature extractors. After pre-processing with YOLOv4, classification was performed with the pretrained VGG-16 model. Accuracy, sensitivity, specificity, precision, F1 score, area under the curve (AUC), and Cohen's kappa performance results of 90.5%, 96.0%, 85.7%, 85.7%, 90.5%, 0.94, and 0.81 were obtained, respectively. Feature extraction was performed with VGG-16 model, and with this extracted features dataset, a 0.9% improvement in accuracy was achieved with the stacking method. Two novel contributions were made to the literature as a result of work performed in this thesis. First, segmental search property was added to the majority voting classifier, providing a 1% improvement in accuracy compared to VGG-16. Second, ANOVA and variance threshold methods were applied to the dataset for feature selection, and the machine learning classifier was inserted into the ANOVA algorithm to find the optimum number of features. The optimal feature set was searched iteratively. With this proposed ANOVA and variance threshold feature selection methods, a 2-5% improvement in performance metrics was achieved with Random Forest, Logistic Regression and Support Vector Machines algorithms. Training time was shortened with both proposed methods.