RESOURCE-EFFICIENT DEEP LEARNING FOR CLINICAL-GRADE VOLUMETRIC SEGMENTATION: THE EFFICIENTMED3D STUDY

Abstract

Deep learning models for 3D medical image segmentation achieve near-inter-rater accuracy but demand 12–24 GB GPU memory and 30–60 s per volume—resources unavailable in most clinical settings. This paper summarizes EfficientMed3D, a resource-efficient framework that couples a depthwise-separable 3D encoder–decoder (3.1M parameters; 2.7 GFLOPs per 128³ patch) with a hardware-aware FP16 inference pipeline (50%-overlap sliding window, Gaussian blending, double-buffered I/O). Benchmarked on BraTS 2021, LiTS 2017, and MSD-Pancreas (1,565 train / 370 test volumes, 8 targets), EfficientMed3D reached a mean Dice of 0.810 versus 0.815 for nnU-Net (−0.5%, below the 1% clinical threshold), while cutting peak memory 58.4% (12.4→5.2 GB), accelerating inference 6.1× (41.3→6.8 s/volume), and reducing energy per volume 79.5%. Ablations attribute 65.6% of latency gains to architecture, 22.3% to overlap tuning, 8.8% to FP16. A TensorRT build on Jetson AGX Orin segmented a volume in 22.4 s—the first sub-minute result on this device class. Efficient, clinical-grade volumetric segmentation is feasible on commodity hardware