Design of an Efficient Bioinspired Multidomain Feature Processing Model for Identification of Spinal Cord Tumours via Convolutional Networks

Classifying spinal cord tumours from Magnetic Resonance Imaging (MRI) is a multidomain task, which involves segmentation of tumour regions, representation of these regions into feature vectors, design of region-specific classifiers, and post-processing techniques. Existing tumour detection models either showcase low-segmentation efficiency, or have higher complexity when evaluated on differential-position MRI scans. Moreover, the accuracy of these models reduces w.r.t. number of disease classes. To overcome these issues, this text proposes design of an efficient bioinspired multidomain feature processing model for classification of spinal cord tumours via convolutional networks. The proposed model initially segments spinal cord images via masked-region augmented Convolutional Neural Network (MRA CNN), which assists in efficient identification of tumour-specific regions. These regions are represented into multidomain features via frequency analysis, Gabor analysis, entropy analysis, Wavelet analysis, and convolution analysis. These feature sets are processed via a Whale Optimization Algorithm (WOA), which assists in identification of highly variant feature sets, which are classified into ‘tumour’ or ‘non-tumour’ types via a Binary CNN (BCNN) process. This process segregates feature vectors into groups of binary classes by maintaining an epitome common class across these sets. The BCNN uses a majority function to identify tumours, which improves the accuracy of classification by 3.9%, precision by 2.5%, recall by 2.8%, and reduce delay by 4.5% when compared with existing tumour detection techniques.