Deep Learning Based Cancer Classification Using Selected Gene Expression Data

Cancer incidence has yearly surged by 23.6 million in 2019, with 10 million deaths from cancer reported in the same year, signifying a 26% and 21% upsurge in cases and deaths over the last decade. Gene expression (GE) analysis is a robust technique for the early diagnosis and classification of cancers by identifying unique molecular characteristics in diverse organs. Unlike other methods, identifying cancer-related genes from GE profiles results in more effective and tailored therapies. This technique, however, suffers from shortcomings such as huge amounts of GE data available, making data extraction and analysis a challenging task, a dearth of inclusive databases, and poor access to GE data extraction technology in some regions. This research proposes a deep learning (DL)-based minimum-redundancy-maximum-relevance (mRMR) technique for the precise identification of six cancer types using GE data, where DL-mRMR merges DL and feature selection (FS) to pick effective features and reduce dataset dimensions. Entrenched in FS, DL-mRMR downsizes the volume of input features reflecting similarities between cancers and within samples of a given tumor, and eventually balances the samples in the database, thereby diagnosing cancer types using a DL classifier. TCGA database simulations reveal a 99% accuracy (ACC) of DL-mRMR in differentiating breast, colorectal, kidney, and lung cancers based on GE data. In addition, due to the reduction in the size of the input sample vector, the complexity of the classifier algorithm and the number of samples required for its training are lower than similar approaches.