DoubleBlock-ViT: A MaxViT-based enhancement and dual-path skip connections for brain tumor segmentation in MRI scans.

Authors

Abstract

Manual interpretation of brain tumor regions in MRI scans demands substantial medical expertise, is time-consuming, and is prone to human error, especially when radiologists must review large scan volumes, leading to fatigue and reduced diagnostic accuracy. This study aims to propose a lightweight segmentation framework with strong global contextual awareness, enabling accurate and efficient delineation of tumor regions for radiological analysis and downstream clinical applications. This study leverages multimodal MRI inputs (T1, T1Gd, T2, T2FLAIR) and introduces a novel hybrid 3D U-Net architecture that integrates Convolutional Neural Networks (CNNs) with Transformer-based modeling. The architecture incorporates a 3D DoubleBlock-ViT Transformer encoder to capture long-range dependencies and global context via attention mechanisms, while a Dual-Path Fusion Block with CNN-based skip connections preserves fine-grained spatial details and enhances feature transfer. Evaluations on BraTS2020 and BraTS2021 benchmark datasets yield Dice scores of 80.11% (ET), 86.60% (TC), and 91.20% (WT) on BraTS2020, and 87.82% (ET), 91.61% (TC), and 92.31% (WT) on BraTS2021, outperforming several state-of-the-art methods. The proposed model delivers state-of-the-art accuracy with only 7.8 million parameters and low computational demands, making it well-suited for clinical deployment and integration into radiomics pipelines for precision oncology. By leveraging multimodal MRI inputs and advanced feature extraction mechanisms, our approach directly aligns with current advances in AI-driven radiomics. The codes and trained models will be publicly available at https://github.com/Laptq201/DoubleBlock-ViT-Unet-segment.

Tags

Topics