Self-supervised 3D Skeleton Completion for Vascular Structures

Jiaxiang Ren¹, Zhenghong Li¹, Wensheng Cheng¹, Zhilin Zou¹, Kicheon Park², Yingtian Pan², Haibin Ling¹

¹Department of Computer Science, ²Department of Biomedical Engineering

Stony Brook University

---

This repository is the official PyTorch implementation of 3D skeleton completion model.

Environment

pytorch == 1.12.1
scikit-image == 0.19.3
napari == 0.4.16
skan == 0.10.0

Inference

Ensure the trained model weight is in weights/model_weights_best.pth. Then run the notebook inference_odt.ipynb for inference.

Datasets and Annotations

• MSD dataset can be downloaded from the official site. The annotation is in dataset/MSD_annotation_3D.csv. Note that only a part of original CT volumes contain vessels so we crop the volumes using the script in dataset/task008_hvessel_preprocessing.ipynb. The annotation is based on the cropped CT patches.

• 10 ODT patches with [ODT image, mask, and skeleton] in dataset/odt_testing.npy are provided for 3D visualization. The whole ODT volomes used in this work is not available for now due to data policy.

Contact

[email protected]

Acknowledgment

This work was partially supported by NIH grants 1R21DA057699, 1RF1DA048808 and 2R01DA029718, and partially supported by NSF grants 2006665 and 2128350.