Official PyTorch implementation of $P^{2}SAM$, from the following paper:

Part-aware Personalized Segment Anything Model for Patient-Specific Segmentation.
Chenhui Zhao and Liyue Shen
University of Michigan
[arXiv]

We propose $P^{2}SAM$, a training-free method to adapt the segmentation model to any new patients relying only on one-shot patient-specific data. $P^{2}SAM$ comprises a novel part-aware prompt mechanism and a distribution-based retrieval approach to filter outlier prompts. These two components effectively mitigate ambiguity and enhance the robust generalization capacity.

• Personalized Segmentation code on PerSeg Dataset
• Fine-tuned Model on NSCLC-Radiomics Dataset
• Patient-Specific Segmentation Code on 4D-Lung Dataset
• Fine-tuned Model on Kvasir-SEG Dataset
• Patient-Specific Segmentation Code on CVC-ClinicDB Dataset
• SAM Fine-tuning Code

See PerSAM to prepare the PerSeg dataset and SAM model weight. We give an example command for personalized segmentation on PerSeg:

First

python p2sam_perseg.py --data '/data/perseg' --outdir '/p2sam_perseg' \
--ckpt '/segment_anything_model/sam_vit_h.pth' --sam-type 'vit_h'\
--min-num-pos 1 --max-num-pos 5 \

Then

python eval_miou_perseg.py --pred-path '/p2sam_perseg' --gt-path '/data/perseg/Annotations' \

This should give

* mIoU 95.6

• For evaluating other model variants, change --ckpt, and --sam-type accordingly.
• Setting --vis for visualization.

Coming soon.

This repository is built using the PerSAM repositories.

If you find this repository helpful, please consider citing:

@article{zhao2024part,
  title={Part-aware Personalized Segment Anything Model for Patient-Specific Segmentation},
  author={Zhao, Chenhui and Shen, Liyue},
  journal={arXiv preprint arXiv:2403.05433},
  year={2024}
}