Graph Based Lung Tumor Segmentation in PET-CT Images

ROS148

Graph Based Lung Tumor Segmentation in PET-CT Images

Scientific Posters

Presented on December 1, 2014
Presented as part of ROS-MOA: Radiation Oncology Monday Poster Discussions

Wei Ju MS, Presenter: Nothing to Disclose

Dehui Xiang BMBS, PhD, Abstract Co-Author: Nothing to Disclose

Zhenxin Wang, Abstract Co-Author: Nothing to Disclose

Bin Zhang, Abstract Co-Author: Nothing to Disclose

Xinjian Chen PhD, Abstract Co-Author: Nothing to Disclose

The aim of this work was to co-segment lung tumor by making use of metabolic information of PET and anatomical information of CT, formulating the segmentation problem as an energy minimization problem based on the graph cut theory.

18 sets of PET-CT images from 18 patients with non-small cell lung cancer (NSCLC) were used to evaluate the efficiency and accuracy of the method. And the Dice similarity coefficient (DSC) and Hausdorff Distance (HD) were used to evaluate the segmentation performance of the proposed method. The method consists of two main steps. First, the pre-processing step was proceeded which includes up-sampling PET images, applying affine registration to PET and CT, and labeling the object and background seeds manually. Then, the graph cut method was applied to segment lung tumor in PET-CT images. The building graph includes two sub-graphs and a special link, in which one sub-graph is for PET and another is for CT, and the special link is a context term which penalize the difference of the tumor segmentation on the two modalities. The cost functions for PET and CT is designed separately. For PET, a novel monotonic downhill cost is proposed which is based on the analysis of the homogeneity and heterogeneity of PET FDG uptake, and a shape penalty cost is also integrated into the cost function which helps to constrain the tumor location during the segmentation. For CT, besides the traditional data and boundary terms, the cost function also includes the shape penalty term which is also used to constrain the tumor location.

The results show that the proposed method has much better segmentation accuracy (on PET, average DSC = 82%, average HD =3.52; on CT, average DSC = 77%, average HD =4.47) compared to the graph cuts methods solely using the PET or CT (p<0.05 ).

We developed a semi-automated graph based method to segment tumor simultaneously on PET and CT image, enabling us to obtain two contours on PET and CT which provides more reliable information for clinical therapists. The quantitative analysis results show that the significant improvement was achieved for tumor delineation.