Related Experiment Video
Updated: Sep 10, 2025
12:08
From Voxels to Knowledge: A Practical Guide to the Segmentation of Complex Electron Microscopy 3D-Data
Published on: August 13, 2014
24.7K
FastSAM3D: An Efficient Segment Anything Model for 3D Volumetric Medical Images
Yiqing Shen1, Jingxing Li1, Xinyuan Shao1
1Johns Hopkins University, Baltimore, MD 21218, USA.
Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention
|August 27, 2025
Summary
FastSAM3D significantly accelerates 3D medical image segmentation using Segment Anything Models (SAMs). This novel approach enables rapid, interactive segmentation of 3D volumes, overcoming previous computational limitations.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- Segment Anything Models (SAMs) offer powerful zero-shot generalization for object segmentation.
- Interactive segmentation with SAMs is valuable for 3D medical imaging but hindered by slow inference and high memory demands.
- Current 2D SAMs applied to 3D volumes cause redundant computations, while 3D SAMs are computationally expensive.
Purpose of the Study:
- To develop a highly efficient method for interactive 3D medical image segmentation using SAMs.
- To overcome the computational bottlenecks of existing SAM approaches for 3D volumetric data.
- To enable real-time, low-cost interactive segmentation on standard GPU hardware.
Main Methods:
- Introduced FastSAM3D, a system for accelerated SAM inference on 3D medical images.
- Employed a layer-wise progressive distillation scheme to transfer knowledge from a large ViT-B to a smaller ViT-Tiny encoder.
- Integrated novel 3D sparse flash attention to reduce memory usage and enhance parallelization.
Main Results:
- Achieved inference speeds of 8 milliseconds for 128x128x128 3D volumes on an NVIDIA A100 GPU.
- Demonstrated speedups of 527.38× over 2D SAMs and 8.75× over 3D SAMs on diverse datasets.
- Maintained segmentation performance without significant decline compared to existing methods.
Conclusions:
- FastSAM3D enables truly interactive, low-cost 3D medical image segmentation with SAMs.
- The developed techniques address critical speed and memory constraints for practical applications.
- This work facilitates the broader adoption of SAMs in clinical 3D medical imaging workflows.
