Table of Contents
Motivation
Method
Results
Cite
Motivation
Using CT images, we can see some unusual structural changes or abnormalities in the lungs. Lobe-wise assessment of these abnormalities is very important for planning the treatment or monitoring the progression. For example, assuming someone has emphysema shown in the CT images as low-attenuation regions, we can detect emphysema by applying some thresholds (LAA-950HU). Once we can automatically segment lobes, we can then compute a per-lobe emphysema percentage, which will help determine whether the patient has heterogeneous emphysema or homogenous emphysema.
Method
Pulmonary lobe segmentation requires the model to recognize local details (voxels around fissures) and capture the global context (structural information over the entire lung). Therefore, I use a coarse-to-fine design based on two cascading U-Nets. At each U-Net, I add a relational neural module to extend the receptive field of the network to be global. The overview of my algorithm can be viewed in the following figure.
please refer to the published paper for more details.
Results
-
We evaluated this algorithm from the COPDGene study (n=1000) and the RadboudCOVID dataset (n=100). In the form of mean (standard deviation), the overall intersection over union (IoU) is 0.949 (0.026) for the COPDGene dataset and 0.922 (0.040) for the RadboudCOVID dataset. The average symmetric surface distance (ASSD) is 0.600 (0.537) for the COPDGene dataset and 0.866 (0.729) for the RadboudCOVID dataset.
-
The externel evaluation was done on Lola11 grand-challenge. The metric for evaluation is a slack version of intersection over union (sIoU), in which they tolarent small amount of errors computed by interobserver disagreement. We achieved overall sIoU at 0.9197 for lobe segmentation and sIoU at 0.9706 for lung segmentation, ranking at 3rd on all automated lobe segmentation algorithms, and 25th of all lung segmentation algorithms.
Cite
If you find this algorithm useful in your research, please consider citing:
@article{Xie20,
author={Xie, Weiyi and Jacobs, Colin and Charbonnier, Jean-Paul and van Ginneken, Bram},
title={Relational Modeling for Robust and Efficient Pulmonary Lobe Segmentation in {CT} Scans},
doi = {10.1109/TMI.2020.2995108},
issue = {8},
volume = {39},
pages = {2664--2675},
journal={IEEE Transactions on Medical Imaging},
year = {2020},
}