Brain Tumor Classification from MRI Scans using a Custom 4-Residual Deep Learning Architecture and Particle Swarm Optimization

Computer aided diagnostic (CAD) are used by radiologists and medical specialists to classify brain tumors from machine generated magnetic resonance image (MRI). Acquisition mechanisms of these images reduces quality of images and therefore sometimes leads to misclassification of the brain tumor. This problem arises a need to develop techniques that can classify brain tumors of different types (i.e. glioma, meningioma and pituitary) with better accuracy. A new automated architecture is proposed in this work for brain tumor classification. A statistical method is employed first to enhance the contrast, then sparse autoencoder based data augmentation is performed. A novel 4RDL-DCNN deep learning network is proposed and trained on the selected MRI dataset. Features are extracted from the global average pool layer of the designed model and further optimized using a nature inspired particle swarm optimization (PSO) algorithm. A famous Figshare brain tumor MRI dataset (publically available) is used to conduct experiments. The obtained accuracy of 98.60% is achieved on this dataset that is improved, then the compared to techniques in the literature.