Reliable Computed Tomography (CT) segmentation is a critical requirement for quantitative imaging and computer-aided clinical workflows. Although Residual U-Net (Res-U-Net) architectures achieve strong overlap performance on curated datasets, threshold-based binarization of probability maps often produces scattered false positives, particularly in low-contrast regions and near complex anatomical boundaries. This study analyses the role of predictive uncertainty in improving structural reliability of CT segmentation outputs. Monte-Carlo dropout is employed at inference time to estimate pixel-wise predictive variance, which is combined with mean probability and component size information within a connected-component framework. A component-level scoring rule is evaluated to suppress unstable, low-confidence regions while preserving coherent anatomical structures. Quantitative experiments demonstrate that uncertainty-aware filtering substantially reduces region-level false positives per scan and improves boundary stability, while maintaining competitive Dice and Intersection over Union (IoU) scores. An ablation study further shows that uncertainty penalization is the primary driver of false-positive reduction, and that combining uncertainty with mild size regularization yields the most balanced performance. The results support the use of uncertainty-guided refinement as a practical reliability layer for Res-U-Net–based CT segmentation systems.