MRI brain tumour scans present significant challenges in medical image processing, particularly in handling DICOM images. DICOM is the standard format for storing medical images, but its complexity creates obstacles in efficiently processing and interpreting the data. DICOM images often contain metadata and multidimensional data, which must be preprocessed correctly to ensure accurate segmentation and classification. Additionally, the inherent variability in brain tumour appearance across patients, coupled with noise and low contrast in MRI scans, makes it difficult to detect tumours with high precision. Segmentation of brain tumours from MRI scans is further complicated by the presence of irregular shapes, varying tumour sizes, and overlapping intensities with healthy tissues. These challenges limit the effectiveness of traditional models in identifying tumour boundaries and classifying tumour regions accurately. Many existing approaches struggle to generalize across different datasets or fail to adapt to new data sources, hindering their practical applicability in clinical settings. To address these limitations, the proposed Att-ConvGLUNet model offers a novel solution for brain tumour segmentation and classification. Proposed Att-ConvGLUNet integrates fully connected dense layers, and attention mechanism combined with ConvGLU activation function, significantly improving performance on the challenging Brats2020, Brats2021 and Brats2023 Dataset. The model achieves impressive metrics, including a 98% Trevesky loss, 97% Dice coefficient. In comparisons with state-of-the-art models like 3D-Unet, Nested Unet, and 3DLadderNet, proposed Att-ConvGLUNet demonstrates superior accuracy, highlighting its potential as a robust and adaptable tool for medical image processing in brain tumour detection.