Handling noisy labels in large-scale datasets remains a significant challenge in machine learning, as manual correction is often impractical. To address this, we introduce OMDMix, a novel framework that leverages semi-supervised learning (SSL) to effectively manage noisy labels. Our approach consists of two primary components: the One-Miss-Drop (OMD) algorithm, which identifies and selects clean samples for an initial labeled dataset, and the Dynamic Sample Selection (DSS) algorithm, which iteratively expands the labeled dataset by integrating additional high-quality labels. By partitioning the dataset into labeled and unlabeled subsets and applying SSL techniques to both, OMDMix enhances model training. Experimental evaluations on CIFAR-10, CIFAR-100, and Clothing1M datasets reveal that OMDMix achieves superior or comparable performance to existing state-of-the-art methods for noisy label scenarios.