GMANet: a CPU-only ultra-lightweight 3D CNN for real-time glioma triage in resource-limited environments
Caijian Hua, Xuerong Jing, Liuying Li, X Y Zhou
Objective Unlike our previously published Graphics Processing Unit-based (GPU) Dynamic Pathway-Enhanced Attention Network (DPEA-Net) which is optimized for radiotherapy planning, state-of-the-art GPU-dependent brain tumor segmentation models cannot be deployed in primary care clinics, community hospitals, or intraoperative Magnetic Resonance Imaging (MRI) settings where dedicated GPU hardware is unavailable. This hardware barrier leads to delayed detection, missed urgent referrals, and prolonged triage in resource-limited environments. Existing lightweight models still suffer from slow Central Processing Unit (CPU) inference and inadequate boundary quality for rapid clinical decision-making. Methods We propose Gated Multi-scale Attention Network (GMANet), an ultra-lightweight 0.20M-parameter 3D CNN explicitly engineered for CPU-only real-time glioma triage. To achieve extreme deployability under strict computational constraints, we introduce three synergistic modules: a gated multi-scale dilated (GMSD) block for efficient multi-scale feature extraction, a full-dimension attention (FDA) module for 3D anatomical continuity, and a learnable transposed upsampling (LTU) module for CPU-optimized boundary recovery. The moderate reduction in enhancing tumor (ET) precision represents a deliberate trade-off to prioritize rapid whole tumor (WT) and tumor core (TC) assessment for initial screening. Results On BraTS2019 and BraTS2021, GMANet achieves WT Dice of 92.45% and 90.93%, TC Dice of 86.28% and 84.72%, and ET Dice of 78.42% and 76.68%, respectively, with only 20.10 Giga Floating-Point Operations (GFLOP). On a standard laptop CPU, inference time is 8.49 s per volume.The ET Dice exceeds the commonly adopted threshold for brain tumor preliminary screening in research practice. This performance renders the model potentially applicable to preliminary triage, pending further prospective clinical validation. Conclusion GMANet trades marginal ET precision for extreme CPU deployability and sub-10-second inference, acting as a potential complementary tool to our GPU-based DPEA-Net. GMANet is designed to support a two-stage clinical pathway: fast CPU screening followed by GPU precision assessment, which requires further implementation and clinical evaluation. Importantly, GMANet is not an incremental improvement of our previously published GPU-based DPEA-Net but a complementary model targeting a distinct clinical scenario where GPUs are unavailable.