CS-MEAN: Cerebral Small Vessel Disease Classification via Deep Mobile Efficient Attention Network

Abstract

Cerebral small vessel disease (CSVD) affects the small blood vessels, that lead to reduced blood flow, stroke, and cognitive decline. However, many existing approaches focus mainly on MRI-based classification and lack effective detection of the disease. To overcome this issue, a novel deep learning (DL)-based CS-MEAN model is proposed for CSVD classification using fundus images from RFMiD dataset. Contrast Limited Adaptive Histogram Equalization (CLAHE) with Frangi (CLAFra) filter is used to enhance the fundus image and to remove the noise in the fundus image. TransU Net contains an encoder-decoder and transformer block to accurately segment the microaneurysm region from the fundus images. The hybrid MobileNet with Efficient Attention (MobileEffi) network extracts the spatial and structural (S2) features gathered from the pre-processed and the segmented images. These S2 features are further fused and classified using the Fully Connected (FC) layer as normal or abnormal. The evaluation metrics of the proposed CS-MEAN are assessed with accuracy, precision, recall, specificity, F1-score, Dice index and Jaccard index. From the experimental analysis, the proposed CS-MEAN model attains an overall accuracy of 99.02% and F1-score of 97.23% for efficient CSVD classification. Moreover, the proposed CS-MEAN model improves the overall accuracy by 1.63%, 1.71%, 0.99%, 0.81% and 0.38% better than SVM, AutoGluon, Mask R-CNN, U-Net with cross-sequence attention, and CNN respectively.