Development of computer aided diagnosis system based on deep learning and ensemble learning methods: Application on omics technologies

Aim: In this study, it is aimed to design a pipeline system, by using open access experimental metabolomics data set on colorectal cancer disease, which can classify the related disease by various ensemble learning and deep learning models and to develop a high-throughput decision support system. Material and Method: In this thesis, the data set produced within the scope of the project numbered PR000226 carried out at Washington University, Department of Anesthesiology and Algology, Northwest Metabolomics Research Center was used. The related data set consisted of a total of 158 samples, including two groups of subjects, 66 CRC patients and 92 healthy controls. As variable selection methods, LASSO, Elastic-Net, Boruta and BorutaShap methods were used. In the classification task, XGBoost, LightGBM, deep neural networks and stacked autoencoder models were utilized. Results: When the findings were examined, it was seen that the model with the worst classification performance is the stacked autoencoder. It was seen that the model cannot achieve the desired classification performance in any variable selection scenario. The LightGBM model gave the best results for all performance measures in the classification of both training and test datasets. In addition, the LightGBM model has achieved this classification performance on the basis of all variable selection methods. Conclusion: In this thesis study, it was seen that ensemble learning methods had much better classification results compared to deep learning methods in all variable selection scenarios. It can be suggested that the use of graphics processing unit (GPU) supported versions of these ensemble learning methods will be more efficient in terms of processing and time costs so that they can provide faster results in large data sets.