Accurate segmentation of cardiac substructures on computed tomography (CT) scans is essential for radiotherapy planning but typically requires large annotated datasets and often generalizes poorly across imaging protocols and patient variations. This study evaluated whether pretrained transformers enable data-efficient training using a fixed architecture with balanced curriculum learning. A hybrid pretrained transformer-convolutional network (SMIT) was fine-tuned on lung cancer patients (Cohort I, N $=$ 180) imaged in the supine position and validated on 60 held-out Cohort I patients and 65 breast cancer patients (Cohort II) imaged in both supine and prone positions. Two configurations were evaluated: SMIT-Balanced (32 contrast-enhanced CTs and 32 non-contrast CTs) and SMIT-Oracle (180 CTs). Performance was compared with nnU-Net and TotalSegmentator. Segmentation accuracy was assessed primarily using the 95th percentile Hausdorff distance (HD95), with radiation dose and overlap-based metrics evaluated as secondary endpoints. SMIT-Balanced achieved accuracy comparable to SMIT-Oracle despite using 64$\%$ fewer training scans. On Cohort I, HD95 was 6.6 $\pm$ 4.3 mm versus 5.4 $\pm$ 2.6 mm, and on Cohort II, 10.0 $\pm$ 9.4 mm versus 9.4 $\pm$ 9.8 mm, respectively, demonstrating robustness to patient, imaging, and data variations. Radiation dose metrics derived from SMIT segmentations were equivalent to those from manual delineations. Although nnU-Net improved over the publicly trained TotalSegmentator, it showed reduced cross-domain robustness compared to SMIT. Balanced curriculum training reduced labeled data requirements without compromising accuracy relative to the oracle model and maintained robustness across patient and imaging variations. Pretraining reduced dependence on data domain and obviated the need for data-specific architectural reconfiguration required by nnU-Net.