Development of a deep learning-based medical decision support system for the diagnosis of pathological voices

The disruption of normal speech flow due to pathological conditions is known as a voice disorder. Therefore, any existing disorder disrupts the speech production system's functioning and produces a distorted voice. Since some laryngeal pathologies are life-threatening, the early detection of voice disorders is important. For this purpose, there is a need to develop a decision support system in the detection of pathological voices. In recent years, machine learning methods have become an interesting research topic to determine pathological voices in order to obtain an individual-level answer, since statistical methods give a group-based result in the evaluation of features extracted from voices. However, since machine learning requires manual extraction of features, deep learning techniques, in which optimal features can be extracted automatically, have become one of the current research topics. However, there are only few research studies on the use of deep learning techniques in the detection of pathological voice disorders. In this thesis study, deep learning methods were used to identify pathological voices. The voice recordings of patients with pathologies causing organic dysphonia due to structural changes in the vocal cords were selected from the Saarbruecken Voice Database. These pathologies included laryngitis, leukoplakia, Reinke's edema, recurrent laryngeal nerve paralysis, vocal cord carcinoma, and vocal cord polyps. The sustained vowel /a/ at the neutral pitch of each individual was selected. The sample included a total of 760 recordings, of which 380 belonged to healthy voices and 380 belonged to pathological voices. The data were divided into training and test sets containing 75% and 25% of the samples, respectively. In the analysis of the samples, first, wavelet noise denoising was applied to the voice signals. Then, the spectrogram images of the voice signals were taken and utilized as inputs in four different Convolutional Neural Network (CNN) architectures, namely GoogleNet, ResNet-50, AlexNet, and SqueezeNet. The selected CNN architectures were used with their own classifiers to determine the pathological voices. Subsequently, the same architectures were employed only for feature extraction, then the Neighborhood Component Analysis employed for feature selection. The performance analyses of the models were undertaken by classifying the selected features with the following classification algorithms: Decision trees, Support Vector Machines, k-Nearest Neighborhood, Ensemble, and a decision tree method designed for this study. The most successful performance was obtained from the method in which the features had been extracted by the SqueezeNet architecture and classified with the Ensemble algorithm. According to the results, the proposed model is promising for the identification of pathological voices.