Related Experiment Video
Updated: Sep 11, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Revisiting model scaling with a U-net benchmark for 3D medical image segmentation
Ziyan Huang1,2,3, Jin Ye3,4, Haoyu Wang1,2,3
1Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, 200240, China.
Scientific Reports
|August 14, 2025
Summary
Larger 3D U-Net models are not always better for medical image segmentation. Optimal architectures depend on the specific task and dataset, with smaller models often performing competitively.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- The 3D U-Net architecture is widely adopted for 3D medical image segmentation.
- The assumption that larger models yield superior performance is prevalent but underexplored.
Purpose of the Study:
- To systematically benchmark U-Net variants for 3D medical image segmentation.
- To challenge the 'bigger is better' paradigm by investigating model size and performance.
- To provide guidelines for optimizing U-Net architectures based on dataset characteristics.
Main Methods:
- Benchmarking 18 U-Net variants with varied resolution stages, depth, and width.
- Evaluating models across 42 diverse public 3D medical imaging datasets.
- Analyzing the impact of architectural choices on segmentation performance.
Main Results:
- Optimal U-Net architectures are task-specific, not universally larger.
- Smaller models frequently achieve competitive performance.
- Increasing resolution stages, depth, or width yields diminishing returns depending on dataset properties.
Conclusions:
- The 'bigger is better' approach is not always optimal for 3D medical image segmentation.
- Dataset characteristics (voxel spacing, anatomical complexity, number of classes) dictate optimal architecture.
- A framework is established for balancing performance and computational efficiency in model selection.
