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

jiaxren@cs.stonybrook.edu

Acknowledgment