Posttraining Network Compression for 3D Medical Image Segmentation: Reducing Computational Efforts via Tucker Decomposition.

Abstract

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To investigate whether the computational effort of 3D CT-based multiorgan segmentation with TotalSegmentator can be reduced via Tucker decomposition-based network compression. Materials and Methods In this retrospective study, Tucker decomposition was applied to the convolutional kernels of the TotalSegmentator model, an nnU-Net model trained on a comprehensive CT dataset for automatic segmentation of 117 anatomic structures. The proposed approach reduced the floating-point operations (FLOPs) and memory required during inference, offering an adjustable trade-off between computational efficiency and segmentation quality. This study utilized the publicly available TotalSegmentator dataset containing 1228 segmented CTs and a test subset of 89 CTs, employing various downsampling factors to explore the relationship between model size, inference speed, and segmentation accuracy, evaluated using the Dice score. Results The application of Tucker decomposition to the TotalSegmentator model substantially reduced the model parameters and FLOPs across various compression ratios, with limited loss in segmentation accuracy. Up to 88% of the model's parameters were removed, with no evidence of differences in performance compared with the original model for 113 of 117 classes after fine-tuning. Practical benefits varied across different graphics processing unit architectures, with more distinct speed-ups on less powerful hardware. Conclusion The study demonstrates that posthoc network compression via Tucker decomposition presents a viable strategy for reducing the computational demand of medical image segmentation models without substantially impacting model accuracy. ©RSNA, 2025.