Spatially Constrained Attention-weighted Fuzzy C-means with Polynomial Hunter-prey Optimization for Brain Tumor Segmentation and Classification

Abstract

Brain tumors refer to the growth of abnormal cells in the human brain, posing a significant risk to a person’s health and well-being. Therefore, early detection of brain tumors is essential for improving diagnosis and providing appropriate treatment. However, existing feature selection models fail to capture the intricate details of brain tumors, resulting in poor classification performance. This study aims to identify accurate tumor boundaries and learn fine-grained details about the brain tumor from the segmented region to enhance brain tumor classification using Magnetic Resonance Imaging (MRI) data. This research proposes a Spatially Constrained Attention Weighted Fuzzy C-Means (SCAW-FCM)-based tumor segmentation model by integrating spatial context and attention weight to achieve more accurate boundary detection. Brain tumor features exhibit complex relationships among textures, shapes, and intensities in MRI. Therefore, a Polynomial Differential Learning Strategy- based Hunter Prey Optimization (PDLS-HPO) algorithm is employed to capture these nonlinear relationships effectively. Furthermore, a Leaky Rectified Linear Unit (ReLU) activation function is incorporated into a Bidirectional Long Short-Term Memory (BiLSTM) model to address the vanishing gradient issue, which helps reduce misclassification by learning fine-grained details in the MRI images, resulting in accurate classification. Experimental results of the SCAW-FCM with PDLS-HPO model demonstrate an accuracy of 99.56% on the Brats 2020 dataset, which is 2.45% higher than existing approaches, such as the Attention-guided Residual Multiscale convolutional neural Network (ARM-Net). The proposed SCAW-FCM-based segmentation model efficiently segments tumor regions and addresses morphological heterogeneity. In addition, the proposed PDLS-HPO-based feature selection model enables the BiLSTM model to identify subtle and complex patterns, thereby efficiently addressing feature redundancy issues. The proposed model improved tumor type classification by learning subtle MRI features, such as tumor texture and boundary characteristics, which are crucial for accurate classification.