Multi-label classification of chest radiographs presents unique challenges due to noisy labels, severe class imbalance, and co-occurring pathologies. In this study, MS-CXR is proposed, a soft-voting ensemble that combines convolutional and transformer-based architectures to improve robustness and generalization. The ensemble integrates four diverse models: a DenseNet-121 trained with MixUp and CutMix, a Swin Transformer, a CoAtNet, and a Vision Transformer pretrained via self-supervised learning on radiographic data (ViT-DINOv2). Evaluation on the ChestX-ray14 dataset under a patient-level split shows that MS-CXR achieves a new state-of-the-art macro-average AUROC of 86.37%, outperforming prior methods including CheXNet and CvTGNet. Per-class performance analysis reveals consistent gains across most pathologies. A weighted-voting variant optimized with SLSQP converges to uniform weights, which reinforces the effectiveness of the base ensemble. These results demonstrate the power of model complementarity and domain-specific self-supervised pretraining for advancing chest X-ray diagnosis.