DEFORMABLE ATTENTION-GUIDED TRANSFORMER FOR BRAIN TUMOR SEGMENTATION

Abstract

Correct segmentation of brain tumors using multimodal magnetic resonance imaging (MRI) is critical for quantitative tumor evaluation, therapy planning and disease monitoring. However, the variability of tumor size, shape, location and appearance among MRI modalities poses a challenge for automatic segmentation. Heterogeneous structure of brain tumors and uneven boundaries of different subregions of a tumor also often impede the effectiveness of traditional segmentation methods. Hence, there is a need for strong approaches that can model local spatial features and long-range contextual dependencies. In this paper, we propose a deformable attention-guided Transformer architecture for multimodal brain tumor segmentation. The framework uses complementary information from T1, T1c, T2 and FLAIR MRI modalities and fuses multi-scale feature extraction, adaptive query refinement, deformable attention and skip connections to improve heterogeneous tumor region representation. We performed experiments on the dataset from the brain tumor image segmentation benchmark (BRATS) multimodal. The model was assessed in terms of Dice coefficient, Precision, Recall and F1-score on Whole Tumor (WT), Tumor Core (TC) and Enhancing Tumor (ET) regions. The suggested model attained Dice scores of 95.8%, 90.7%, and 86.1% for WT, TC and ET, respectively, while the average Dice score was 90.9%. Ablation analysis further highlighted the contribution of deformable attention, query refinement, skip connections and multimodal feature integration to the segmentation performance. The model also showed a good trade-off between segmentation accuracy and computational efficiency, which suggests the possible use of this approach in automated MRI processing workflows. The experimental results illustrate the potential of the proposed framework for accurate and robust multimodal MRI based brain tumor segmentation.