Mean-enabled Laplacian U-Net for efficient glioma sub-region segmentation in multimodal MRI
TL;DR
A deep learning-assisted, low-complexity architecture termed mean-enabled Laplacian U-Net (MELU-Net) is proposed for glioma sub-region segmentation, which improves boundary representation and feature consistency while maintaining lower computational complexity compared to attention-based and transformer-based models.
Abstract
Multimodal magnetic resonance imaging (MRI) is essential for accurate delineation of glioma sub-regions, including background (BG), enhancing tumor (ET), tumor core (TC) and edema (ED). However, heterogeneous intensity distributions and indistinct tumor boundaries pose significant challenges for precise segmentation. Although deep learning methods have improved performance, many existing models require high computational resources and lack efficiency for practical deployment. In this paper, a deep learning-assisted, low-complexity architecture termed mean-enabled Laplacian U-Net (MELU-Net) is proposed for glioma sub-region segmentation. The model extends the conventional U-Net by integrating channel-wise mean feature aggregation with Laplacian-based edge enhancement within skip connections. This design improves boundary representation and feature consistency while maintaining lower computational complexity compared to attention-based and transformer-based models, making it suitable for resource-constrained environments. Resource-constrained environments refer to computing platforms with limited computational power, memory capacity, storage, and energy availability, such as embedded systems, edge AI devices and mobile healthcare platforms. Therefore, MELU-Net is designed as a lightweight architecture suitable for deployment in such settings. The proposed method is evaluated on the BraTS 2021 dataset using both single and stacked multimodal MRI inputs through qualitative and quantitative analyses. For single-modality input, the model achieves Dice scores of 0.9943 (BG), 0.4824 (ET), 0.4208 (TC) and 0.6012 (ED), demonstrating improved segmentation over the baseline. For stacked multimodal input, significant performance gains are observed, with Dice scores of 0.9966 (BG), 0.7529 (ET), 0.6555 (TC) and 0.7910 (ED). These results highlight the effectiveness of multimodal feature fusion in ET sub-region delineation. Overall, MELU-Net provides a robust and computationally efficient framework for accurate glioma segmentation, making it suitable for real-world clinical applications.