Bridging the AI Hardware Gap in Resource Constrained Settings: Lightweight Architectures and Weight Optimization for 3D Medical Image Segmentation with nnU-Net

State-of-the-art 3D medical image segmentation frameworks like nnU-Net deliver exceptional performance but typically require high-end GPU infrastructure, creat-ing severe bottlenecks for clinical deployment in resource-constrained settings such as Low- and Middle-Income Countries (LMICs). To bridge this hardware gap, we present a practical, end-to-end framework integrating initialization-time Jacobian-sensitivity pruning (XTinyU-Net), post-training INT8 quantization, and CPU-specific weight redistribution within the nnU-Net v2 pipeline. Using three open-source benchmark datasets of varying complexity—the ACDC cardiac MRI dataset, the BraTS-Africa glioma dataset, and the BraTS-METS dataset—we demonstrate that sensitivity-guided pruning converges on a compact architecture with 1.03 million parameters, reducing the baseline model size 13.3-fold from 0.8 GB to 0.06 MB. By shifting intensive training to cloud-based GPUs while executing optimized, quantized inference on standard local CPUs, our approach achieves dramatic performance gains. Specifically, combining an INT8-quantized OpenVINO backend with the disabled test-time augmentation (TTA) pipeline yields speed-ups of 48 to 95 times on CPU, bringing inference times close to native GPU performance. Crucially, a paired Wilcoxon signed-rank test with Bonferroni correction reveals that this compressed CPU pipeline matches or significantly exceeds the segmentation accuracy (Dice and surface Dice) of the uncompressed GPU baseline across nearly all evaluation structures, without suffering statistically significant performance drops. This lightweight strategy successfully eliminates local GPU dependency, offering a viable, highly efficient solution for deploying advanced artificial intelligence models in resource-limited clinical environments.