Simultaneous brain tumor segmentation and glioma classification with MRI are critical in clinical practice for precise diagnostics, treatment planning, and decision-making. Despite significant efforts, robust results remain difficult to achieve because of tumor heterogeneity, tumor boundary irregularity, and cross-modal differences between datasets. In this work, we design a hybrid attention transformer network for brain tumor segmentation and HGG/LGG classification with multi-modal MRIs. Specifically, our proposed network employs the Pyramid Vision Transformer v2 (PVT-v2) backbone along with a hybrid decoder consisting of a convolutional upsampling path, channel-spatial hybrid attention blocks, and a transformer. Our proposed method leverages multi-modal MRIs. We trained our model on the BraTS 2018 dataset and validated its performance using the BraTS 2019 dataset to ensure robust and reliable evaluation. The proposed framework demonstrates a competitive level of accuracy in both segmentation and classification tasks, highlighting its effectiveness and reliability, while still being computationally efficient. The proposed framework was trained for 100 epochs for the final performance evaluation, whereas baseline comparisons and ablation experiments were conducted over 50 epochs. The suggested model outperforms numerous leading-edge models with 16,354,057 parameters, At 50 epochs, the proposed framework attained a Dice Similarity Coefficient (DSC) of 0.8005 and an AUC of 0.8965, while at 100 epochs, it attained a DSC of 0.8352 ± 0.1031 for segmentation and an AUC of 0.9385 for classification.

Segmentation, classification, Brain Tumor, Transformer, Deep Learning, HGG, LGG