DCLA-UNet: dynamic cross-layer compressive attention and large axial separable convolution for multi-modal MRI brain tumor segmentation

Abstract

Accurate brain tumor segmentation from multi-modal Magnetic Resonance Imaging (MRI) is critical for clinical prognosis, yet existing 3D architectures face a persistent dilemma. Traditional Convolutional Neural Network (CNN) suffer from restricted receptive fields, while emerging Vision Transformers incur prohibitive quadratic computational complexity and excessive parameterization, severely hindering their deployment on resource-constrained clinical devices. To address this fundamental trade-off between global contextual modeling and computational efficiency, this study proposes DCLA-UNet, a highly parameter-efficient 3D segmentation framework driven by two core architectural innovations. First, we design a lightweight encoder-decoder structure integrating a Slim Large-Kernel Module (SLKM) that utilizes depthwise axial convolutions to enlarge the effective receptive field, coupled with a Multi-Scale Fusion Module (MSFM) employing parallel dilated branches for robust semantic reconstruction. Second, we introduce a Dynamic Cross-Layer Compressive Attention (DCLA) mechanism that leverages extreme single-channel compression and multi-kernel spatial alignment to adaptively modulate bottleneck features, effectively bridging the semantic gap and suppressing background noise. Extensive evaluations across the BraTS (2019, 2020, 2021) and MSD BrainTumour bench- marks systematically validate the efficacy and generalization capability of DCLA-UNet. The proposed architecture requires only 0.53M trainable parameters—a 94.76% reduction compared to the 3D U-Net baseline—and consumes just 39.23 Giga Floating Point Operations (GFLOPs). Despite this extreme compression, DCLA-UNet achieves a competitive mean Dice Similarity Coefficient (DSC) of 87.5 ± 0.6% on BraTS2021, outperforming emerging lightweight architectures such as SegFormer3D and Mamba3D. While it matches the DSC of heavier models like MogaNet using only 9.2% of its parameters, a trade-off is observed in the 95th percentile Hausdorff Distance (HD_95) (7.6 ± 1.8), objectively reflecting the inherent limits of boundary precision under strict parameter constraints. DCLA-UNet successfully breaks the conventional bottleneck of 3D global context modeling by achieving a competitive balance between parameter efficiency and segmentation accuracy. It provides a robust, parameter-efficient solution for 3D medical image segmentation in storage-constrained clinical scenarios. (The source code is available at https://github.com/Helium-327/DCLA-UNet-3D ).