PAM: A Propagation-Based Model for Segmenting Any 3D Objects across Multi-Modal Medical Images

Paper | Checkpoint | Tutorial-PAM | Tutorial-PAM-for-2D | Tutorial-PAM-for-Pathology

PAM (Propagation-based Anything Model) is a general-purpose framework that produces volumetric 3D segmentations from a minimal 2D prompt . It works across multi-modal medical images including CT, MRI, PET, SRX, micro-CT, and more—without retraining.

Features

• Minimal → Volumetric Segmentation : A single 2D prompt slice is enough for PAM to propagate and generate a complete 3D segmentation.
• Multi-modal & Object-agnostic : Trained across 44 diverse datasets including CT, MRI, PET, and SRX.
• No Fine-tuning Required : A fully reusable workflow—upload your volume → provide one prompt → obtain full volumetric segmentation, with no retraining needed.
• Fast, Robust, and User-efficient : Reduces interaction time and inference cost; stable across prompt and propagation variations; particularly strong for irregular and challenging anatomical structures.

Download Pretrained Weights

PAM provides pretrained weights that can be downloaded from:

• Baidu Pan（百度网盘）

  Link: https://pan.baidu.com/s/1WdTPG_ojCEeWR5qpl2_WzQ?pwd=mksh

  Code: mksh

• Google Drive

  Link: https://drive.google.com/drive/folders/10fi5ZLFRnW5lCMVI7vMY61eTfOQlN9oI?usp=share_link

After downloading, place all checkpoint files into:

tutorials/checkpoints/

Tutorials

PAM supports multiple forms of prompt-driven semantic propagation. These tutorials cover (1) the core 3D propagation pipeline reported in the paper and (2) two extended capabilities discovered after publication that greatly expand PAM's usability.

1. Prompt-to-Volume Propagation (Core Workflow)

tutorials/3d-propagation.ipynb

Reproduces PAM's primary capability described in the paper. Given a single 2D prompt slice, PAM propagates semantics across the entire volume to produce a full 3D segmentation. This works across CT, MRI, PET, SRX, micro-CT, and more — without retraining.

2. Cross-2D Image Propagation (Extended Capability)

tutorials/cross-2d-image-propagation.ipynb

A practical extension discovered after paper acceptance. PAM's strong semantic consistency allows it to propagate annotations horizontally across large sets of 2D medical images. With only a few annotated images, PAM can generate high-quality pseudo-labels at scale, enabling efficient dataset construction.

3. Whole-Slide Pathology Propagation (Extended Capability)

tutorials/pathology-propagation.ipynb

Demonstrates PAM's surprising generalization to computational pathology. From a few annotated instances (e.g., one or a few cells), PAM can propagate semantics across an entire whole-slide image (WSI), locating all corresponding structures despite never being trained on pathology. This enables scalable annotation of dense structures in gigapixel WSIs.

Datasets Used in PAM (Links)

We will provide a complete table of dataset access links used in PAM here, which include updated URLs compared with the Supplementary Table 1 in the supplementary materials.

ID	Dataset	Download link	Modality	Objects
D01	AbdomenCT-1K	link	CT	Kidneys, liver, pancreas, spleen
D02	AMOS-CT	link	CT	Aorta, bladder, duodenum, esophagus, gallbladder, left kidney, liver, left adrenal gland, prostate/uterus, pancreas, postcava, right kidney, right adrenal gland, spleen, stomach
D03	AutoPET-PETCT	link	PET-CT	Lesion
D04	AutoPET-CT	link	CT	Lesion
D05	COVID-19 Seg. Challenge	link	CT	COVID-19 infections
D06	COVID-19-CT-Seg	link	CT	COVID-19 infections, left lung, right lung
D07	HECKTOR	link	PET-CT	Head & neck lymph nodes, head & neck primary tumor
D08	INSTANCE	link	CT	Hematoma
D09	KiPA	link	CT	Kidney tumor, kidney, renal artery, renal vein
D10	KiTS	link	CT	Kidney cyst, kidney tumor, kidney
D11	Lymph nodes	link	CT	Mediastinal lymph node
D12	NSCLC Pleural Effusion	link	CT	Effusions, thoracic cavities
D13	MSD-Task03 Liver	link	CT	Liver, liver cancer
D14	MSD-Task06 Lung	link	CT	Lung cancer
D15	MSD-Task07 Pancreas	link	CT	Pancreas cancer, pancreas
D16	MSD-Task08 HepaticVessel	link	CT	Hepatic tumor, hepatic vessel
D17	MSD-Task09 Spleen	link	CT	Spleen
D18	MSD-Task10 Colon	link	CT	Colon cancer primaries
D19	Total Segmentator	link	CT	Full multi-organ segmentation
D20	AMOS-MR	link	MR	Multi-organ
D21	ATLAS-R2.0	link	MR-T1	Brain stroke
D22	BraTS	link	Multi-MR	Enhancing tumor, edema, non-enhancing core
D23	ISLES	link	MR-DWI, MR-ADC, MR-FLAIR	Ischemic stroke
D24	MnM2	link	MR	LV, myocardium, RV
D25	NCI-ISBI	link	MR-ADC, MR-T2	Prostate zones
D26	PI-CAI	link	MR-bp	Prostate cancer
D27	PROMISE	link	MR-T2	Prostate
D28	Qin-Prostate-Repeatability	link	MR	Prostate, suspected tumor
D29	Spine	link	MR	Sacral, lumbar, thoracic spine
D30	MSD-Task01 BrainTumour	link	MR	Edema, enhancing tumor, non-enhancing tumor
D31	MSD-Task02 Heart	link	MR	Left atrium
D32	MSD-Task04 Hippocampus	link	MR	Anterior + posterior hippocampus
D33	MSD-Task05 Prostate	link	MR	Peripheral zone, transitional zone
D34	WMH	link	MR-T1, MR-FLAIR	WMH lesions
D35	Adrenal-ACC-Ki67-Seg	link	CT	Adrenocortical carcinoma
D36	CHAOS-CT	link	CT	Liver
D37	HaN-Seg	link	CT	Head & neck structures
D38	HCC-TACE-Seg	link	CT	Liver blood vessels, liver tumor
D39	LNQ2023	link	CT	Mediastinal lymph node
D40	QUBIQ	link	CT	Pancreas, pancreatic lesion
D41	WORD	link	CT	Multi-organ
D42	ACDC	link	MR	LV, myocardium, RV
D43	CHAOS-MR	link	MR-T1, MR-T2	Kidney, liver, spleen
D44	MouseKidney-SRX	link	SRX	Glomerulus

Citation

If you find PAM useful, please cite:

@article{chen2025pam,
  title={PAM: A propagation-based model for segmenting any 3d objects across multi-modal medical images},
  author={Chen, Zifan and Nan, Xinyu and Li, Jiazheng and Zhao, Jie and Li, Haifeng and Lin, Ziling and Li, Haoshen and Chen, Heyun and Liu, Yiting and Tang, Lei and Li, Zhang and Bin, Dong},
  journal={npj Digital Medicine},
  year={2025},
  volume={8},
  number={753},
  doi={10.1038/s41746-025-02087-y},
  url={https://doi.org/10.1038/s41746-025-02087-y}
}