Positron emission tomography (PET) plays an essential role in lung cancer diagnosis, staging, and treatment. However, it is difficult to accurately segment and separate tumors residing in close proximity. It is even more challenging for tumor segmentation from PET due to its heterogeneous density distribution and the difficulty in finding the stopping criterion for delineation. To address these issues, in this paper, we investigated the tumor segmentation and separation by using Tumor-Customized Downhill (TCD) method and compared TCD with other widely used methods including 40% and 50% of maximum SUV, and watershed technique. Our quantitative and qualitative comparison and validation on seven clinical studies, including thirteen tumors demonstrated that TCD outperformed its counterpart methods in terms of tumor segmentation and separation.