Multi-label chest X-ray classification faces challenges such as significant variations in lesion scale, underutilized latent correlations among diseases, and poor model robustness caused by clinical label noise interference. To this end, this study proposes a chest X-ray disease classification network named MS-SPNet based on the multi-scale shared prototypical network and confidence guidance. Firstly, a multi-scale attention front-end module is introduced, which employs DenseNet-169 as the backbone network and adaptively fuses features at different levels to precisely capture lesion features across varying scales, ranging from tiny nodules to large-scale effusions. Secondly, an affinity-guided shared prototype mechanism is proposed, which leverages prior knowledge of disease co-occurrence to guide the construction of a shared prototype pool. It models anatomical correlations between diseases while reducing feature representation redundancy. Finally, a prototype-guided confidence weighting strategy is designed to mitigate the influence of poor annotation quality. This strategy dynamically evaluates sample reliability and adjusts loss weights based on the stability of prototype activation, thereby effectively suppressing the interference of noisy labels on the training process. Experimental results demonstrate that the proposed method achieves an average AUC of 85.1% on the ChestX-ray14 dataset. Furthermore, it yields an average AUC of 84.7% on the larger-scale CheXpert dataset (under the U-Zeros setting). The two metrics are significantly superior to those of existing state-of-the-art algorithms, thus fully verifying that the model exhibits excellent generalization performance while maintaining its intuitive interpretability.