FFCAEs: An efficient feature fusion framework using cascaded autoencoders for the identification of gliomas

Abstract

AbstractIntracranial tumors arise from constituents of the brain and its meninges. Glioblastoma (GBM) is the most common adult primary intracranial neoplasm and is categorized as high‐grade astrocytoma according to the World Health Organization (WHO). The survival rate for 5 and 10 years after diagnosis is under 10%, contributing to its grave prognosis. Early detection of GBM enables early intervention, prognostication, and treatment monitoring. Computer‐aided diagnostics (CAD) is a computerized process that helps to differentiate between GBM and low‐grade gliomas (LGG), using the perceptible analysis of magnetic resonance (MR) of the brain. This study proposes a framework consisting of a feature fusion algorithm with cascaded autoencoders (CAEs), referred to as FFCAEs. Here we utilized two CAEs and extracted the relevant features from multiple CAEs. Inspired by the existing work on fusion algorithms, the obtained features are then fused by using a novel fusion algorithm. Finally, the resultant fused features are classified with the Softmax classifier to arrive at an average classification accuracy of 96.7%, which is 2.45% more than the previously best‐performing model. The method is shown to be efficacious thus, it can be useful as a utility program for doctors.