Swinme: Swin Transformer V2-based Framework for Multimodal Brain Tumor Segmentation

Abstract

Accurate and efficient brain tumor segmentation from Magnetic Resonance Imaging (MRI) is a critical component in neuro-oncology, as it directly supports diagnosis, treatment planning, and disease progression assessment. This requirement is particularly significant in pediatric gliomas, where early and precise clinical intervention can substantially influence patient outcomes. Despite its importance, manual tumor delineation remains a time-consuming and subjective process, highly dependent on expert availability and prone to inter-observer variability. These limitations restrict scalability and reliability, especially when dealing with high-resolution three-dimensional MRI data and in healthcare environments with limited resources. Consequently, there is a strong need for automated segmentation methods that are both accurate and computationally efficient. This thesis proposes a novel hybrid deep learning framework, termed SwinME-UNETR 3D, for automated brain tumor segmentation. The proposed approach integrates the hierarchical attention mechanism of Swin Transformer V2 with a multi-scale enhancement strategy (SwinME) and the volumetric modeling capability of UNETR architecture. The framework is designed to capture long-range spatial dependencies while preserving fine-grained local features that are essential for precise tumor boundary delineation. Multi-scale feature enhancement and an enhanced transformer module are incorporated to strengthen feature representation across different spatial resolutions, enabling effective handling of tumor heterogeneity and complex anatomical structures present in three-dimensional MRI volumes. The proposed method is evaluated using the BraTS 2023 dataset, which comprises multimodal MRI scans including T1, T1 post-contrast (T1ce), T2, and FLAIR sequences. Segmentation performance is assessed on clinically relevant tumor subregions, namely whole tumor (WT), tumor core (TC), and enhancing tumor (ET), using Dice Similarity Coefficient and sensitivity metrics. Experimental results demonstrate robust and consistent performance, achieving an average Dice score of 0.91 on the validation set. x Qualitative analysis further confirms strong alignment with expert annotations and high volumetric consistency. These findings indicate that transformer-based architectures augmented with multi-scale enhancement provide an effective solution for automated brain tumor segmentation and hold strong potential for future clinical decision-support applications.