Detection and Classification of Automated Brain Stroke Lesion with optimized Dual Stage Deep Stacked Auto-Encoder

Abstract

Brain Stroke is defined as the sudden death of the brain cells due to lack of blood circulation and form a lesion/mass in the cerebral parenchyma and led to loss of speech, weakness, or paralysis of one side of the body. If the diseases are detected in early stage it will be cured. The existing method does not provide efficient accuracy. In this paper two type of brain stroke lesions are classified such as infarct (lack of blood supply) and Haemorrhagic stroke (breaking of blood vessel). In this manuscript, Automated Brian Stroke Lesion Detection and Classification using Non-Contrast Computed Tomography and Dual Stage Deep stacked auto-encoder (DS-DSAE) with an Evolved Gradient Descent optimisation (EGDO) method is proposed to detect the brain stroke in early stage with great accuracy. In this the input image are taken from the slice level of Non-Contrast CT images dataset. Then the images are pre-processed, images are enhanced by removing skull regions, then the rotations are performed by mid-line symmetry process. Then the ROI region is extracted using wavelet domain. Then the images are classified using DS-DSAE and the weight parameters of the DS-DSAE are tuned using EGDO algorithm. Then the abnormal portions of the brain stroke lesions are detected and classified as acute infarct, chronic infarct and ischemic stroke, haemorrhagic stroke, and normal. The objective function is to increase the accuracy by decreasing the computational complexity.  The simulation process is executed in the MATLAB platform.  The proposed CLACHE-IDFNN-MBO attains higher accuracy 99.56%, High Precision 88.74%, High F-Score 92.5%, High Sensitivity 94.23%, High  Specificity 91.45%, lower computational time 0.019(s) and the proposed method is compared with the existing methods such as Fractional Order BAT Algorithm Fuzzy C with Delaunay triangulation (DT), social group optimization (SGO) and Fuzzy-Tsallis entropy (FTE), moth-flame algorithm (MFOA) and Kapur’s thresholding  respectively.