Related Experiment Video
Updated: Jan 9, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
PAM: a propagation-based model for segmenting any 3D objects across multi-modal medical images
Zifan Chen1,2, Xinyu Nan2, Jiazheng Li3
1Center for Machine Learning Research, Peking University, Beijing, China.
NPJ Digital Medicine
|December 2, 2025
Summary
We developed PAM, a novel framework for 3D medical image segmentation. This efficient tool generates accurate segmentations from minimal 2D prompts, outperforming existing methods and reducing manual annotation needs.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Volumetric segmentation in medical imaging is challenging due to extensive annotation requirements.
- Current methods lack transferability across different objects and require task-specific retraining.
Purpose of the Study:
- To present PAM, a propagation-based framework for generating 3D segmentations from minimal 2D prompts.
- To improve the efficiency and generalizability of automated medical image segmentation.
Main Methods:
- PAM integrates a CNN-based UNet for intra-slice feature extraction.
- Transformer attention is employed for inter-slice propagation, capturing structural and semantic continuity.
- The framework was evaluated across 44 diverse medical imaging datasets.
Main Results:
- PAM significantly outperformed MedSAM and SegVol, with a 19.3% average improvement in Dice Similarity Coefficient (DSC).
- Stable performance was observed across prompt and propagation variations.
- Inference time was reduced, and user interaction time decreased by 63.6%.
Conclusions:
- PAM provides accurate 3D segmentations from minimal input, reducing reliance on manual annotation and retraining.
- The framework demonstrates robust cross-object generalization, particularly for irregular shapes.
- PAM offers an efficient and generalizable tool for automated clinical imaging.
