Insertion Network

Insertion Network is a method for building slice-level correspondence between two axial body CT image sequences.
Given a test CT volume and a template CT volume, the model establishes correspondence by inserting slices from the template image into their correct locations within the test image sequence.

---

Overview

• Task: Slice-level correspondence between two axial CT volumes
• Input:
  • Test CT image (NIfTI format)
  • Template CT image (NIfTI format)
• Output: Predicted insertion locations of template slices within the test sequence
• Core idea: Correspondence is built by predicting where each template slice should be inserted into the ordered slice sequence of the test image.

---

Installation

Install the required dependencies:

pip install -r requirements.txt

---

Inference

To run inference and compute slice correspondence between two CT volumes:

python inference_args.py \
  --test_nii <test image path> \
  --template_nii <query image path>

The output is a visualization showing the predicted insertion positions of the template slices within the test sequence.

Example Output (query image referes to the template image)

---

Training

The training code is provided at:

exp/Exp_body_ct_semi_supervised_1.0_kl_sinPE_1/train.py

This implementation can be used as a reference for training a new model on your own dataset.

Notes on Training

• The provided training setup is supervised.
• Slice-level ground-truth correspondence is required.
• Please refer to the paper for details on how the ground-truth correspondences are constructed.

---

Citation

If you use this code or the Insertion Network method in your work, please cite:

@inproceedings{su2026insertion,
  title={Insertion Network for Image Sequence Correspondence Building},
  author={Su, Dingjie and Hong, Weixiang and Dawant, Benoit M. and Landman, Bennett A.},
  booktitle={SPIE Medical Imaging: Image Processing},
  year={2026}
}