Modified Recurrent Residual Attention U-Net model for MRI-based brain tumor segmentation

Brain tumors area leading cause of neurological impairment, often resulting in severe consequences, including fatality. Timely detection of brain tumors is imperative for effective intervention, with Magnetic Resonance Imaging (MRI) standing out as the most efficient method for identifying irregularities in the brain. In this paper, we introduce a novel approach for MRI-based brain tumor segmentation (BTS): The Modified Recurrent Residual Attention U-Net (Mod-R2AU-Net). The proposed architecture integrates recurrent and residual components within both encoding and decoding paths, leveraging Recurrent Residual Convolutional Layers to dynamically capture temporal dependencies and subtle image features. Attention gates (AGs) are employed to enhance feature refinement and segmentation accuracy, replacing conventional skip connections. Before segmentation, MRI images undergo a sophisticated preprocessing phase aimed at optimizing image fidelity, including noise reduction and spatial alignment to ensure the highest quality input for the segmentation model. The Mod-R2AU-Net model outperforms existing segmentation models, notably due to the integration of AGs resulting in improved segmentation accuracy. Performance evaluation includes objective metrics such as Binary Accuracy, Dice coefficient, and Intersection over Union, along with a subjective assessment of the segmentation results. Additionally, our research extends to seamlessly integrating the model into a comprehensive pipeline designed to refine BTS in medical images, enhancing precision in detecting and delineating intricate brain tumor boundaries and advancing the effectiveness of medical image analysis.