Topology-Preserving Automatic Labeling of Coronary Arteries via Anatomy-aware Connection Classifier
Zhixing Zhang*, Ziwei Zhao*, Dong Wang, Shishuang Zhao, Yuhang Liu, Jia Liu, Liwei Wang
[7/22/2023] [MICCAI 2023] [Arxiv]
In this repo, we provide labels for the open-source dataset used in the paper "Topology-Preserving Automatic Labeling of Coronary Arteries via Anatomy-aware Connection Classifier".
Given a vessel tree, we can split it into segments using bifurcation rules illustrated in fig.1. That is, we use consecutive bifurcation points to mark the boundary of one vessel segment. Each of the vessel segment has only ONE class.
In our setting, we only consider the centerlines of vessels. Therefore, we only store point sets of vessel centerlines in our label files, and our goal is to classify each of the *centerline segments. The labeling of vessels can be simply done by a knn mapping.
The labels are stored in .json files, and each json has the following structure:
{
l_points=[{point_set=[[...]], cls=c},...],
r_points=[{point_set=[[...]], cls=c'},...],
l_tree=[[x,y],...],
r_tree=[[x',y'],...]
}l_points: A list containing centerline point coordinates and the vessel segment labels of l_tree. Each element of the list is a dictionary with 2 fields: point_set (all centerline point coordinates of this vessel segment of the shape N*3) and cls (label of this segment).
r_points: A list containing centerline point coordinates and the vessel segment labels of r_tree. Each element of the list is a dictionary with 2 fields: point_set (all centerline point coordinates of this vessel segment of the shape N*3) and cls (label of this segment).
l_tree: A list storing the structure of l_tree, each element [x, y] is an edge of the tree, where x is the x-th segment in l_points and y is the y-th segment in l_points.
r_tree: A list storing the structure of r_tree, each element [x, y] is an edge of the tree, where x is the x-th segment in r_points and y is the y-th segment in r_points.
You need to first download the original open-source dataset.
Then, you need to clone this repo, and unzip labels.zip to see the labels. The password to unzip the .zip file can be found at here with code ga35. After unzipping the files, you need to align the label files with the images. The file name of the labels are the same as the corresponding images. Note that you probably need to reverse (and/or) swap the images' axis to align them. Before training your model, you should visualize the images and labels to ensure the alignment is correct.
Here we provide a sample code for alignment:
import os, numpy, json
def load_json(file_path):
with open(file_path, 'r') as f:
data = json.load(f)
return data
def load_image(file_path):
...
return image
img=IMG_PATH
label=LABEL_PATH
out_path=OUT_PATH
os.makedirs(out_path, exist_ok=True)
f=os.listdir(img)
# Somehow filter out irrelevent files in f
for f_ in f:
fid = f_.split('.')[0]
image = load_img(os.path.join(img, f_))
h, w, d = image.shape
lb = load_json(os.path.join(label, fid+'.json'))
for s in lb['l_points']:
point=s['point_set']
# reverse w axis for example
point[:, 1] = w - point[:, 1]
for s in lb['r_points']:
point=s['point_set']
# reverse w axis for example
point[:, 1] = w - point[:, 1]
# you can adjust spacing here
...
np.savez_compressed(os.path.join(out_path, fid+'.npz'), img=image, label=lb)If you have any question, feel free to contact Zhixing Zhang at zhangzhixing@stu.pku.edu.cn .
If you find this repository useful, please consider starring this repo & citing our paper:
@inproceedings{zhang2023topology,
title={Topology-Preserving Automatic Labeling of Coronary Arteries via Anatomy-Aware Connection Classifier},
author={Zhang, Zhixing and Zhao, Ziwei and Wang, Dong and Zhao, Shishuang and Liu, Yuhang and Liu, Jia and Wang, Liwei},
booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
pages={759--769},
year={2023},
organization={Springer}
}