Related Experiment Video
Updated: Jun 29, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Improving medical image segmentation in pre-trained U-Nets using Shapley-guided pruning of adaptive skip-connection
Emil Benedykciuk1, Marcin Denkowski2, Grzegorz Marcin Wójcik1
1Institute of Computer Science and Mathematics, Maria Curie Sklodowska University, Akademicka 9 Street, 20-033, Lublin, Poland.
Scientific Reports
|June 27, 2026
Summary
Shapley-guided pruning enhances neural architecture search (NAS) for medical image segmentation by reducing computational costs and improving stability. This method efficiently refines pre-trained U-Nets on limited data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Neural Architecture Search (NAS) shows promise for medical image segmentation.
- Practical application is hindered by high computational demands and unstable architectures, especially with pre-trained models and limited datasets.
Purpose of the Study:
- To introduce Shapley-guided pruning, a practical extension of retrofit NAS for pre-trained U-Nets.
- To address computational cost and stability issues in NAS for medical image segmentation.
Main Methods:
- Implemented iterative pruning driven by Shapley value estimates on validation data.
- Preserved learned architecture parameters while removing low-impact components.
- Utilized existing IAC search space and PC-DARTS-style supernet optimization.
Main Results:
- Improved or matched strong baselines across four public benchmarks (ACDC, BraTS, KiTS, AMOS).
- Accelerated search time by up to four times in a controlled 2D, single-GPU setting.
- Demonstrated more stable architecture choices across multiple runs.
Conclusions:
- Shapley-guided pruning offers a practical, resource-efficient retrofit-NAS mechanism.
- The method enhances reliability and reduces complexity for medical image segmentation tasks.
- Findings are within a specific 2D slice-based protocol, not directly comparable to high-resolution 3D pipelines.