The gap between cutting-edge medical AI and the clinics that need it most has long been a quiet crisis in neuro-oncology. Glioma triage depends on fast, accurate MRI interpretation — yet the GPU hardware required to run state-of-the-art segmentation models is largely absent from primary care settings, community hospitals, and intraoperative suites. A model capable of running on ordinary CPUs in real time could meaningfully reduce the diagnostic delay that worsens outcomes in high-grade brain tumors.
GMANet (Gated Multi-scale Attention Network) is a 3D convolutional neural network engineered with just 0.20 million parameters — orders of magnitude leaner than conventional segmentation architectures — and designed explicitly for CPU-only deployment. The network combines three functional modules: a gated multi-scale dilated block for extracting features across spatial scales without high memory overhead, a full-dimension attention module to preserve three-dimensional anatomical context, and a learnable transposed upsampling module optimized for boundary reconstruction under computational constraints. Evaluated on the BraTS2019 and BraTS2021 benchmarks, GMANet achieved whole-tumor Dice scores of 92.45% and 90.93% respectively, and strong tumor-core performance — with developers explicitly accepting a modest reduction in enhancing-tumor precision as a deliberate design trade-off favoring rapid triage over radiotherapy-grade delineation.
This work sits at a meaningful intersection of clinical need and engineering pragmatism. Most medical AI benchmarking implicitly assumes GPU availability, which systematically excludes the lower-resource environments where diagnostic delays are most consequential. GMANet's framing as a triage rather than a treatment-planning tool is a refreshingly honest scope limitation — whole-tumor and tumor-core classification are clinically sufficient to trigger urgent referral. Limitations worth noting include validation on benchmark datasets rather than prospective clinical cohorts, and the absence of head-to-head CPU inference timing data against competing lightweight models. Still, the architectural approach — trading segmentation granularity for deployability — represents an incremental but practically important contribution to equitable neuro-oncology care.