This document provides a brief guide for setting up and using PromptReg, our demonstration implementation currently under active development.

Before you start, make sure your environment meets these dependencies:

cuda==11.4
monai==1.3.0
nibabel==5.2.0
torch==1.12.0
torchvision==0.13.0
transformers==4.39.0

We have provided a sample dataset in the .example folder to facilitate a quick start with both intra-subject and inter-subject registration tasks. Here’s how the dataset is structured:

.example/
|--- inter_subject/
| |--- prostate_mr
| | |--- image1.png
| | |--- image2.png
| | |--- image3.png
| | |--- image4.png
| | |--- label1.png // binary label for targeted ROI
| | |--- ...
| |--- abdomen_mr
| | |--- ...
| |--- abdomen_ct
| | |--- ...

|--- intra_subject/
| |--- aerial
| | |--- zh7_01_04_R.png // fixed image #1
| | |--- zh7_01_04_T.png // moving image #1
| | |--- zh7_02_01_R.png // fixed image #2
| | |--- zh7_02_01_T.png // moving image #2
| |--- histology
| | |--- ...
| |--- cardiac_mr/
| | |--- ...
| |--- lung_ct/
| | |--- ...
| |--- cell/
| | |--- ...

In the intra-subject folder, each pair of images is denoted as '_R' for the moving image and '_T' for the fixed image, which are to be aligned.

To use your own data, simply add it to the example folder and specify the image paths when running the demo:

--fix_image ./example/intra_subject/aerial/zh7_01_04_R.png --mov_image ./example/intra_subject/aerial/zh7_01_04_T.png

Easily select a SAM variant by specifying the sam_type in your command:

--sam_type sam_h

Available models include:

• sam_b: Segment Anything (SAM-B)
• sam_h: Segment Anything (SAM-H)
• medsam: Segment Anything in Medical Images (MedSAM)
• slimsam: 0.1% Data Makes Segment Anything Slim (SlimSAM)
• sam_hq: Segment Anything in High Quality (SAM_HQ)

To incorporate pretrained weights, update your model via huggingface.

To run a quick demo for SAMReg [paper], use the following command:

python demo.py 

For interactive use in PromptReg, add a prompt point within the image dimensions:

python demo.py --prompt --prompt_point [80,113]

To perform traditional image warping with interpolation, specify the ROI type:

python demo.py --prompt --interpolate --ROI_type label_ROI 

Select from label_ROI or pseudo_ROI depending on your needs.

For more precise wrapping, adjust the number of interpolation iterations (default: 10,000).

--ddf_max_iter 100000

Set a custom precision threshold for adopting paired ROIs:

--sim_criteria 0.88

Set a custom number of adopted paired ROIs:

--num_pair 5

• If the number of adopted paired ROIs (num_pair) exceeds the number of generated paired ROIs, the code will automatically switch to using the precision threshold (sim_criteria).
• The --sim_criteria argument becomes ineffective when --num_pair is set.

Set a filter for ROIs using the --v_min and --v_max arguments. ROIs with an area less than v_min or greater than v_max will be excluded.

--v_min 100 --v_max 7000

This work was supported by the International Alliance for Cancer Early Detection, a partnership between Cancer Research UK [C28070/A30912; C73666/A31378], Canary Center at Stanford University, the University of Cambridge, OHSU Knight Cancer Institute, University College London and the University of Manchester.

@article{huang2024one,
  title={One registration is worth two segmentations},
  author={Huang, Shiqi and Xu, Tingfa and Shen, Ziyi and Saeed, Shaheer Ullah and Yan, Wen and Barratt, Dean and Hu, Yipeng},
  journal={arXiv preprint arXiv:2405.10879},
  year={2024}
}