Evaluation of methods of co-segmentation on PET/CT images of lung tumor: simulation study

Introduction: Lung cancer is one of the most common causes of cancer-related deaths worldwide. Nowadays PET/CT plays an essential role in radiotherapy planning specially for lung tumors as it provides anatomical and functional information simultaneously that is effective in accurate tumor delineation. The optimal segmentation method has not been introduced yet, however several methods have been proposed up to now. Lake of suitable gold standard for correct evaluating of segmentation algorithms is the most important reason of it. The Monte-Carlo simulations are the ideal tool since it allows a complete description of data not accessible in the case of patient studies. The aim of this study was to simulate clinically realistic Monte-Carlo PET/CT data, obtained using the NCAT numerical phantom and the GATE simulation tool, to compare precision of 3 co- segmentation algorithms on lesion delineation.   Materials and Methods: In this study, detection system and geometry of PET/CT scanner were modeled using GATE simulation toolkit. After validation of simulated scanner and modification of GATE parameters, a digital phantom was designed using NCAT digital phantom based on patient's information extracted from PET/CT images. Then 5 sphere-shape tumors with different sizes were created in lung as lesions. In order to implement co-segmentation algorithms, simulated images of phantoms were merged with CT images by AMIDE (a Medical Image Data Examiner). Finally simulated pulmonary lesions were segmented using 3 co-segmentation matlab codes: Watershed, Region growing and Graph cut. Comparing of these methods was performed by parameters included: Tumor Diameter (TD) and Gross Tumor Volume (TV). At the end, results of PET segmentation and PET/CT segmentation were compared as well in order to evaluate effect of using CT information in precision of tumor volume detection.   Results: TD and GTV measured from PET/CT images segmented by region growing had difference less than 5% from gold standard and this method was efficient for all 5 tumor sizes. Graph cut outcomes had GTV and TD difference less than 15 % and it could not detect tumors with size less than 3 cm. watershed did not have correct outcomes. The co-segmented tumor volume on PET-CT image was most strongly correlated with specified tumor volume of phantoms compared with co-segmented tumor volume on PET individually.   Conclusion: Results demonstrate that among 3 segmentation algorithm, region growing had the most accurate outcomes and watershed was the weakest method. It also showed accuracy of segmentation methods was depended on tumor size especially for graph cut. Tumors of less than 3 cm were indistinguishable by graph cut method and watershed was an incapable algorithm for tumor delineation. Accuracy on GTV and TD measurement using all 3 segmentation methods was largely improved on PET/CT fused image compared to PET image individually as anatomical information of CT.