Deep learning-based cancer classification for microarray data: a systematic review

Abstract

Deep neural networks are robust techniques and recently used extensively for building cancer classification models from different types of data. Nowadays, microarray gene expression datasets consider an essential source of data that is used in cancer classifications. However, due to the small size of samples compared to the high dimensionality of microarray data, many machine learning techniques have failed to distinguish the most relevant and informatics genes. Therefore, deep learning is demand due to its ability to automatically discovering the complex relationship between features with significant accuracy and high performance. The current study aims to reveal the state-of-the-art of deep neural network architectures and how it can utilize from microarray data. Therefore, several deep neural network architectures were built such as CNN, DNN, RNN, DBN, DBM and DAE to be compatible with the different learning processes (supervised, unsupervised and semi-supervised). As a result, CNN considers the most common neural network architecture used in the medical field due to its robustness and high performance in cancer classification. Results indicate that choosing suitable architecture of the deep neural network and its hyperparameters is one of the most difficulties facing the researcher in designing models for cancer prediction and classification because there is no particular rule to ensure high prediction accuracy.