Lightweight multimodal brain tumor MRI segmentation network for efficient clinical deployment: MLG-UNet achieves an average Dice of 86.67% with 0.81 M parameters

To address the problems of high model redundancy, high inference latency, and high computational resource demands faced by brain tumor MRI segmentation models when deployed in radiology departments and clinical settings, this study explores methods to maintain high segmentation accuracy while minimizing model parameters and computational overhead as much as possible, so as to meet the needs of efficient assisted analysis on general clinical terminals and resource-constrained devices.

Based on the BraTS2020 and BraTS2019 datasets from the international multimodal Brain Tumor Segmentation (BraTS) challenge, only the official training cases were used and divided at the patient level into training sets (BraTS2020: 277 cases; BraTS2019: 251 cases), validation sets (BraTS2020: 37 cases; BraTS2019: 34 cases), and test sets (BraTS2020: 55 cases; BraTS2019: 50 cases), after which 2D axial slices were extracted. A lightweight multimodal brain tumor MRI segmentation model, termed MLG-UNet, was proposed, which adopted a shared encoder-decoder architecture and incorporates 3 designed modules: branch-context multi-kernel block (BCMB), low-rank bilinear fusion module (LRBF), and lightweight global bottleneck modulator (LGBM), which enhance multi-scale feature extraction, cross-layer feature interaction, and high-level semantic modeling, respectively. The model was compared with representative methods, including U-Net, TransUNet, UNeXt, EMCAD, Swin-Unet, and model performance and stability were comprehensively evaluated through ablation experiments, visualization analysis, and patient-level five-round repeated holdout validation.

On the BraTS2020 test set, MLG-UNet achieved a mean Dice of 86.67% and a mean 95th-percentile Hausdorff distance (HD95) of 3.61 mm with only 0.81 M parameters and 0.53 G floating-point operations (FLOPs), and its mean metrics were superior to all compared methods. The Dice values for the whole tumor (WT), tumor core (TC), and enhancing tumor (ET) subregions reached 91.50%, 88.82%, and 79.70%, with corresponding HD95 values of 3.37, 3.58, and 3.89 mm, respectively. Compared with TransUNet, which achieved the second-best mean Dice, MLG-UNet improved the mean Dice by 0.49 percentage points, while reducing the parameters from 105.17 M to 0.81 M. Compared with the lightweight method UNeXt, MLG-UNet improved the mean Dice by 0.69 percentage points and reduced the mean HD95 by 0.38 mm, with lower parameter count and computational complexity. On the BraTS2019 test set, the mean Dice was 86.51% and the mean HD95 was 3.68 mm, also outperforming all compared methods, demonstrating favorable cross-dataset generalization. Ablation experiments showed that BCMB, LRBF, and LGBM all effectively improved segmentation performance. Visual comparison demonstrated higher agreement between the segmentation results of the proposed method and expert annotations. In patient-level five-round repeated holdout validation, the mean Dice was (86.71±0.52)% and the mean HD95 was (3.65±0.14) mm, indicating high stability.

MLG-UNet achieved a mean Dice of 86.67%, with only 0.81 M parameters, maintaining high brain tumor MRI segmentation accuracy while controlling model size and computational overhead, and exhibits stable segmentation performance across different patient partitions.