Classification of acute leukaemias with a hybrid use of feature selection algorithms and deep learning-based architectures

Abstract

The microarray technology which is preferred in the fields of medicine and biology is an analysis method that produces quantitative or qualitative data. It has a strong potential for revealing and interpreting patterns between genes. To reveal this potential, it is possible to provide a molecular evaluation of cancer diseases associated with genes. However, microarray datasets have a high dimensional structure. This is known as the curse of dimensionality in machine learning. The main aim is to give a helpful idea to the experts by using computer-aided systems to facilitate the evaluation process on microarray datasets. In this study, a high-dimensional microarray dataset is analyzed for the classification of acute leukaemias. In the first phase of the study, ant colony, whale and particle swarm optimization algorithms are used to select disease-related genes from the dataset. Selected potential genes were evaluated with classical machine learning algorithms. These genes obtained in the second stage of the study were expressed as spectrograms by the wavelet transform method. In the third stage of the study, the CLAHE method is used to improve the local contrast in the spectrograms. Finally, the obtained improved spectrograms are classified by transfer learning architectures and DGCNN (deep graph convolutional neural network) approach. The maximum success rates obtained as a result of the classification of the spectral density information of the selected genes using the ant, particle swarm and whale feature selection algorithms with the DGCNN approach are found to be 93.33%, 86.6% and 86.6%, respectively.