An optimized hierarchical attention assisted deep learning model for brain tissue classification.

BACKGROUND & OBJECTIVE: Precise differentiation of brain tissue from Magnetic Resonance Imaging is a vital constraint in various medical applications. Several studies have conducted on brain tissue classification and segmentation, which offer limitations due to the inconsistency of brain tissues caused by different scanner types and acquisition procedures. Image segmentation is a critical task due to the complex structures of tissues. As a result, this paper developed an effective, optimized hierarchical deep learning method for detecting brain tissue anomalies. METHOD: This research was initiated with pre-processing of MRI images using min-max normalization. Furthermore, a novel hybrid segmentation approach named Residual fused accumulated U-net bridge module (ResFAU-net) is used, which combines residual blocks, attention gates, and the Fused Accumulation Bridge module. After segmentation, the Hierarchical Attention-Based Modified Convolutional Cascaded Capsule Network (HAMC3) is used for classification. COMPARISON WITH EXISTING METHODS: This model integrates the benefits of cascaded capsule networks, convolutional neural networks (CNN), and the hierarchical attention mechanism. To reduce complexity, the parameters are modified using the Coati optimization algorithm. The demanding assessment of model functioning is carried out with the BRATS2020 dataset. RESULTS: Experimental findings of the proposed model on the BRATS2020 dataset are validated with a wide range of performance indicators such as dice score, Intersection over Union (IoU), accuracy, precision, sensitivity, specificity, and F1-score. CONCLUSIONS: The proposed model performs well, with an IoU score of 96.29 % and an accuracy of 99.03 %. This evaluation was mainly used to detect brain tissue abnormalities by segmenting and classifying the abnormal brain tissue region.