Refining skip connections in convolutional encoder–decoder networks for whole meningioma segmentation using shifted window transformer blocks

Abstract

[EN]Accurate segmentation of meningiomas in magnetic resonance imaging scans is essential for clinical planning, yet remains challenging due to their irregular shapes and subtle boundaries. In this study, we refine skip connections in convolutional encoder–decoder networks (widely known through the U-Net architecture) by selectively integrating shifted window transformer blocks. Unlike prior transformer-based architectures, which primarily enhance encoder or decoder stages, our approach targets shallow skip connections to improve the fusion of local detail and global context. An ablation study on the BraTS Meningioma 2023 dataset demonstrates that applying transformer blocks to the first two skip levels yields an optimal balance between accuracy and efficiency. The proposed model achieves a Dice similarity coefficient of 0.9119, outperforming conventional encoder–decoder baselines such as U-Net, Attention U-Net, and a widened U-Net variant, while delivering more precise boundary delineation with competitive recall.