In this paper, we propose an uncertainty-based coarse-to-fine segmentation framework, recognizing that reliable tumor segmentation in chest CT suffers from boundary ambiguity, class imbalance, and anatomical variation. The framework combines full-volume tumor localization with fine segmentation of regions of interest (ROIs), and is enhanced by anatomically-aware postprocessing. A first-stage model generates coarse predictions and performs anatomically-informed filtering based on lung overlap, proximity to the lung surface, and component size. The resulting ROIs are segmented by a second-stage model trained with an uncertainty-aware loss function to improve accuracy and boundary correction in the ambiguous region. Experimental results on private and public datasets demonstrate improved Dice and Hausdorff scores, reduced false positives, and improved spatial interpretability. These results highlight the importance of combining uncertainty modeling and anatomical priors in a cascaded segmentation pipeline to generate robust and clinically meaningful tumor contours. On the Orlando dataset, the proposed framework improves the Dice score of Swin UNETR from 0.4690 to 0.6447, and the reduction in erroneous components is strongly correlated with the improvement in segmentation performance, demonstrating the value of anatomically informed postprocessing.
