Related Experiment Video
Updated: Jun 15, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Post-Training Network Compression for 3D Medical Image Segmentation: Reducing Computational Efforts via Tucker
Tobias Weber1, Jakob Dexl1, David Rügamer1
1From the Department of Radiology, University Hospital, LMU Munich, Marchioninistr 15, 81377 Munich, Germany (T.W., J.D., M.I.); Department of Statistics, LMU Munich, Munich, Germany (T.W., D.R.); and Munich Center for Machine Learning, Munich, Germany (T.W., J.D., D.R., M.I.).
Network compression using Tucker decomposition significantly reduces computational demands for 3D CT segmentation with TotalSegmentator. This deep learning approach achieves substantial parameter reduction with minimal impact on segmentation accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Science
Background:
- Accurate 3D CT-based multiorgan segmentation is crucial for clinical applications.
- Deep learning models like TotalSegmentator offer automated segmentation but require significant computational resources.
- Network compression techniques are explored to enhance efficiency.
Purpose of the Study:
- To evaluate Tucker decomposition for compressing the TotalSegmentator model.
- To reduce computational effort (floating-point operations, memory) during inference.
- To assess the trade-off between compression, inference speed, and segmentation accuracy.
Main Methods:
- Applied Tucker decomposition to convolutional kernels of the TotalSegmentator (nnU-Net) model.
- Utilized a large CT dataset (1228 scans) and a test subset (89 scans).
- Evaluated segmentation performance using Dice scores across various downsampling factors.
Main Results:
- Achieved substantial reduction in model parameters (up to 88.17%) and floating-point operations.
- Demonstrated minimal loss in segmentation accuracy, with no significant difference for 113 of 117 classes post-fine-tuning.
- Observed varied practical benefits across GPU architectures, with speedups on less powerful hardware.
Conclusions:
- Post hoc network compression via Tucker decomposition is a viable strategy for medical image segmentation.
- Reduces computational demand of deep learning models without substantial accuracy compromise.
- Enables more efficient deployment of advanced segmentation tools.
