Accurate lung nodule segmentation is critical for reliable chest CT analysis, yet current models rely heavily on large-scale, manually annotated data and often struggle to generalize to unseen datasets. This study investigates whether the transferability of an existing segmentation model can be improved without additional human-labeled lesion masks. Building on the TotalSegmentator lung-nodule task, a weakly supervised framework is proposed in which model-generated predictions serve both as pseudo-labels and as sources of synthetic anomalies, which are blended into CT volumes via 3D Poisson blending. The approach employs a two-stage DRAEM-inspired architecture, combining a pre-trained reconstruction network that captures normal-appearance priors with a discriminative network trained to localize nodular anomalies. Evaluation on the LNDb dataset shows that the proposed model improves over raw TotalSegmentator predictions on multiple metrics, increasing the Dice score from 23.87 to 26.73. However, post-processed TotalSegmentator masks still achieve higher performance (Dice 32.25), and both approaches exhibit substantial performance degradation, underscoring the challenges posed by domain shift. These findings suggest that while weak supervision with synthetic anomalies can guide feature learning, matching the accuracy and robustness of fully supervised methods remains challenging.