An Optimized Transformer-Driven Approach for Multi-Modal Brain Tumor Segmentation in Neuro-Oncological MRI Imaging

Accurate brain tumor segmentation from multi-modal magnetic resonance imaging (MRI) is a critical task in healthcare, directly supporting early diagnosis, surgical planning, radiotherapy targeting, and longitudinal treatment assessment in neuro-oncology. However, reliable segmentation remains challenging due to complex tumor heterogeneity, irregular anatomical structures, and intensity in homogeneity across MRI modalities, which limit the effectiveness of conventional image processing and standard deep learning (DL) and Artificial Intelligence (AI) approaches. To address these challenges, this study proposes a transparent and healthcare-centric DL framework, termed Swin-UNETR-GJO, which integrates the hierarchical self-attention capability of the Swin Transformer with the encoder–decoder architecture of UNETR for multi-modal brain tumor segmentation. To further enhance model adaptability and performance, the Golden Jackal Optimization (GJO) algorithm is employed to automatically adjust architectural and training parameters, such as attention window size, learning rate, embedding dimensionality, and dropout. The proposed model is trained and evaluated on the BraTS 2021 benchmark dataset. In addition, attention-based activation and feature attribution maps are incorporated. To support clinical interpretability and trust in AI-assisted diagnosis, the two computational explainability measures, Region-Attention Overlap (RAO) and Boundary-Attention Accuracy (BAA), are incorporated, enabling visual delineation of tumor sub-regions. The proposed Swin-UNETR-GJO achieved Dice scores of 0.91 for Whole Tumor (WT), 0.89 for Tumor Core (TC), and 0.87 for Enhancing Tumor (ET), with a minimum Hausdorff Distance 95% (HD95) of 6.1 mm; five-fold cross-validation further confirmed stable performance, with mean Dice Similarity Coefficient (DSC) values of 0.908 ± 0.008, 0.884 ± 0.005, and 0.868 ± 0.008, respectively. The proposed framework provides accurate, stable, and interpretable brain tumor segmentation, highlighting its potential application as a reliable decision-support tool for healthcare.