| It has been reported that incidence and mortality of lung cancer are significantly increasing in recent decades. Therefore, accurate lung tumor segmentation plays a key role in clinical lung cancer evaluation and monitoring. PET-CT, as a quantitative molecular-anatomic modality, now is the best way for clinical cancer early diagnosis.In this study, we make fully use of PET and CT two imaging modalities, combined 3D downhill function, blind source separation method and random forests to solve the lung tumor segmentation problem. Firstly, 3D monotonic downhill function is applied to localize homogeneous lung tumor roughly and then a random forest method with BSS features is proposed to segment lung tumor accurately. In localization phase, homogeneous tumors can be seen as a monotonic decreasing function, that is to say, intensity transition from the maximum value in homogeneous tumor region to the background is gradual. In segmentation phase, features are extracted from PET and CT images simultaneously during random forest training. Also, we proposed a new feature set called BSS features, which reflect decomposition information of lung tumor in PET image. And then, feature selection step is used to reduce the training time and improve the classification performance.In this paper, the novelties include three parts: 1) feature extraction was based on the combination of PET and CT modalities; 2) a new feature set called BSS feature, which take full advantage of the decomposition information in PET image; 3) 3D downhill function and random forest method are fully integrated to improve the accuracy of lung tumor segmentation. The proposed algorithm was validated on a dataset consists of 23 PET-CT images which are from the patients with Non-small cell lung(NSCLC) cancer. The proposed algorithm was compared with region growing, improved graph cut method. The comparative experimental results indicate that our method have an excellent performance in accuracy. |
PDF Full Text Request