Whole Heart Segmentation (WHS) plays a crucial role in the diagnosis and treatment of cardiovascular diseases (CVDs), which remain the leading cause of death worldwide. Despite notable progress in medical image segmentation using deep learning models, the task of WHS continues to pose significant challenges due to anatomical variability, imaging artifacts, and the scarcity of high-quality, diverse 3D datasets. Existing datasets are often limited in size, collected within a single medical institution, and focused on a narrow range of heart structures, resulting in poor model generalization to new domains. In this study, we investigate methods to enhance the reliability and generalizability of deep learning models for WHS. We construct a combined dataset from two publicly available sources, aiming to mitigate the limitations of each by ensuring a more diverse representation of clinical cases and including data from at least two distinct medical domains. To address the complexities of merging heterogeneous datasets, we propose a dedicated preprocessing pipeline, which includes a localization step for key anatomical structures. The model trained on this unified dataset demonstrates strong generalization capabilities and achieves competitive segmentation performance. While it slightly underperforms compared to baseline models trained separately on each dataset, our results highlight the viability of domain-agnostic WHS solutions and the importance of dataset diversity and preprocessing strategies in improving clinical applicability.