IMPROVING COMPUTATIONAL EFFICIENCY OF DEEP LEARNING ALGORITHMS FOR 3D MEDICAL IMAGE SEGMENTATION
Keywords:
3D medical image segmentation; computational efficiency; deep learning; depthwise separable convolution; mixed precision inference; nnU-Net; medical image computingAbstract
Three-dimensional (3D) medical image segmentation underpins computer
aided diagnosis, radiotherapy planning, and image-guided surgery, yet state-of-the-art
volumetric networks impose prohibitive costs: 12–24 GB of GPU memory, 480–900
GFLOPs per patch, and 30–60 s of inference per volume. These requirements exclude
deployment on the commodity hardware available in most clinical settings. In this
study we propose EfficientMed3D, a framework that combines (i) a lightweight
encoder–decoder built from depthwise separable 3D convolutions and multi-scale
residual bottleneck blocks (3.1M parameters, 2.7 GFLOPs per 128³ patch), (ii) a
compound Dice–Focal objective with instance normalization enabling small-batch
training, and (iii) a hardware-aware inference pipeline integrating FP16 mixed
precision, 50%-overlap sliding windows with Gaussian blending, and asynchronous
double-buffering. On BraTS 2021, LiTS 2017, and MSD-Pancreas, EfficientMed3D
achieved a mean Dice score of 0.810 versus 0.815 for nnU-Net (−0.5%, below the
predefined 1% clinical threshold), while reducing peak training memory by 58.4%
(12.4→5.2 GB), accelerating inference 6.1× (41.3→6.8 s/volume), and cutting energy
References
[1] Ronneberger O., Fischer P., Brox T. U-Net: Convolutional networks for
biomedical image segmentation. MICCAI, LNCS 9351, pp. 234–241, 2015.
[2] Çiçek Ö., Abdulkadir A., Lienkamp S.S., Brox T., Ronneberger O. 3D U
Net: Learning dense volumetric segmentation from sparse annotation. MICCAI, pp.
424–432, 2016.
[3] Milletari F., Navab N., Ahmadi S.-A. V-Net: Fully convolutional neural
networks for volumetric medical image segmentation. 3DV, pp. 565–571, 2016.
[4] Isensee F., Jaeger P.F., Kohl S.A.A., Petersen J., Maier-Hein K.H. nnU-Net:
a self-configuring method for deep learning-based biomedical image segmentation.
Nature Methods 18(2):203–211, 2021.
[5] Howard A.G., et al. MobileNets: Efficient convolutional neural networks for
mobile vision applications. arXiv:1704.04861, 2017.
[6] Li X., Chen H., Qi X., Dou Q., Fu C.-W., Heng P.-A. H-DenseUNet: hybrid
densely connected UNet for liver and tumor segmentation from CT volumes. IEEE
TMI 37(12):2663–2674, 2018.
[7] Roth H.R., et al. Spatial aggregation of holistically-nested convolutional
neural networks for automated pancreas localization and segmentation. Medical Image
Analysis 45:94–107, 2018.
[8] Sokolova K., Kuric D., Shalbanov R., Chernyavskiy A., Zykov S. Sparse
annotation training for 3D medical image segmentation. MICCAI DART workshop,
2020.
[9] Micikevicius P., et al. Mixed precision training. ICLR, 2018.
