A hybrid deep learning framework using DT-FLBP and entropy features for stroke detection in MRI images

Abstract

Cerebrovascular diseases such as strokes seriously affect a person's life and good health. The diagnosis and treatment of stroke are significantly aided by the quantitative analysis of the brain using Magnetic Resonance Imaging (MRI) images. The prime intention of this research is to design an effective Hybrid Xception-ShuffleNet (HX-ShuffleNet) for detecting stroke disease. Initially, an MRI image is acquired from the database. Then, the acquired MRI image is fed into the image denoising module, where image denoising is performed using a median filter. Later, the stroke lesion segmentation is done based on the U-Net to isolate the stroke lesions from the entire image. After stroke lesion segmentation, image augmentation (random rotation, shifting, shearing, flipping) is done. Features are extracted using Dual-Tree-Fuzzy Local Binary Pattern (DT-FLBP), which combines Dual-Tree Complex Wavelet Transform (DTCWT), Fuzzy Local Binary Pattern (FLBP), and entropy. For stroke detection, HX-ShuffleNet, a fusion of Xception and ShuffleNet models, is used, achieving a True Positive Rate (TPR) of 0.928, accuracy of 0.935, True Negative Rate (TNR) of 0.929, Precision of 0.922, and F1-score of 0.928.