Bridging global context and local precision using a disagreement-based region specific ensemble of Swin UNETR and...

Brain tumor is one of the most challenging neurological diseases to diagnose and even a minor inaccuracy in the tumor characterization can be fatal. An accurate and reliable brain tumor segmentation from 3D MRI images is a fundamental requirement for an effective diagnosis, treatment planning and assessment of outcome in neuro-oncology. Due to infiltrative growth of tumors, heterogeneity in its structure and diffuse boundaries of tumor regions, brain tumor segmentation is quite critical and challenging. Even a minor error in delineation can adversely affect surgical resection and radiotherapy planning. To address these challenges, this study proposes a region-adaptive ensemble framework that integrates the complementary strengths of two capable 3D segmentation models, SegResNet and Swin UNETR through a staged fusion strategy: simple averaging, region-adaptive soft weighting (RSW), and a disagreement-based region-specific refinement (DRE) for high-conflict voxels. The CNN-based SegResNet is capable in capturing fine-grained local textures and well-defined tumor cores due to its convolutional local bias whereas Transformer-based Swin UNETR is capable in modeling long range contextual dependencies across MRI volume due to its hierarchical Transformer architecture. These two models are finetuned on BraTS 2020 dataset and then integrated using a dynamic voxel-wise disagreement-based fusion strategy that adaptively weights model predictions based on parameters like regional confidence, historical performance and level of disagreement. The multi-run experimental evaluation of the architecture on BraTS 2020 dataset is able to achieve impressive dice scores of 0.9447 ± 0.0021 in Whole Tumor (WT), 0.9231 ± 0.0033 in Tumor Core (TC) and 0.9071 ± 0.0041 in Enhancing Tumor (ET) regions. These results indicate that the Region Adaptive fusion with a disagreement-based refinement between convolutional and transformer-based models leads to a robust framework for brain tumor segmentation, that upon further research and validation might turn out to be suitable for clinical decision making and treatment planning.