Bipolar hastalığı sınıflandırması için fNIRS ölçümleri kullanılarak eğitilmiş derin ağların görselleştirilmesi

Deep learning applications have achieved impressive performances on many medical problems such as classification of disorders, effects of a treatment or unspotted symptoms of a disease, etc. While modern deep learning progress is impressive in such areas, genuine understandings of its working principles are not clear. For that matter, the term black box has often been associated with deep learning algorithms. The majority of previous studies have concentrated on networks' successes and have computed their performances in terms of accuracy levels. However, this thesis focuses on disintegrating the internal working mechanisms of neural networks into intuitive and understandable components. It makes them easy to understand and to interpret from medical experts' perspectives. With this purpose in mind, pre-trained Convolutional Neural Networks and Residual Neural Networks are utilized by using time-series neuroimaging data, i.e. Functional Near-Infrared Spectroscopy (fNIRS) measurements, belonging to two classes, namely healthy and bipolar, and their visualization outputs are attained. Since these outputs are complex time-series data, they are analyzed by statistical methods such as chi-square and t-tests so that the intrinsic features of healthy and bipolar subjects specific to their classes are obtained. Results are compared with previous medical studies and are analyzed so that potential reasons behind the classification results are provided. The contribution of this thesis is providing an inference about visualization outcomes of different neural networks, which are trained for the bipolar disorder classification using fNIRS data. Therefore, this study tries to fill the void between medical researchers and deep learning experts.