Diagnostic oncology requires the use of medical image segmentation, which is especially challenging when analyzing brain tumors due to their complexity and diversity. Several innovative computational methods, including CNN and other deep learning techniques, have successfully addressed certain challenges. This research aimed to improve the accuracy of categorizing and segmenting tumor regions using a U-Net architecture in MRI scans. This enables improved identification and description of tumor subregions, such as necrotic core, edematous areas, augmenting regions, and surrounding healthy tissues, to aid in diagnosis. MBAA-UNet, a groundbreaking deep learning architecture, was developed to enhance multi-modal medical picture segmentation, particularly with the help of brain tumor MRIs. This study is part of this effort. Its suggested architecture comprises a multi-modal feature extraction pathway that extracts large sets of images (up to 90% of all imaging data), an enhanced edge detection mechanism that maps tumor margins with boundary attention, and the integration of extracted features into ’feature fusion and refinement’ modules. By dynamically altering feature relevance, adaptive attention techniques can enhance the segmentation accuracy of complex tumor borders and improve the edge detection ability of the system. Using the BraTS 2023 dataset for experimental validation, the proposed architecture was demonstrated to exhibit high precision in segmenting tumors of various subregional types. In terms of key metrices such as Dice similarity coefficient(DSC), Dice loss, sensitivity and mean IoU, the MBAA-UNet model outperforms the existing models.