Feature selection based on hybrid optimization for magnetic resonance imaging brain tumor classification and segmentation

Abstract

With the health information technology being infused into clinical health, e-health is becoming a key factor in delivering improvements in the health sector. Brain tumor data feature selection is crucial for the development of a viable cancer detection system based on brain tumor data. Our study aimed to obtain an optimal feature subset through a hybrid algorithm of Simulated Annealing-Genetic Algorithms (SA-GA). Two real datasets of brain tumor Magnetic Resonance Images are used to assess the performances of the proposed approach. The first dataset was freely downloaded from the Harvard Medical School brain atlas. The second brain tumor dataset was created from Nanfang Hospital, Guangzhou, China, and General Hospital, Tianjing Medical University, China from 2005 to 2012. The proposed approach is compared to the methods of simulated annealing, genetic algorithm and with the state-of-the-art methods used separately. The obtained results show that SA-GA exceeds simulated annealing and genetic algorithms when they are applied in isolation, in terms of accuracy and computing time. The evaluation shows that our method overtakes the state-of-the-art methods with a segmentation accuracy rate of 97.82%±0.74 for glioma tumor and 95.12% ±3.21for pituitary tumor.