Conventional image segmentation methods typically require precise pixel-level or bounding box annotations. This study seeks to develop a weakly-supervised framework for COVID-19 X-ray segmentation, utilizing text descriptions as training input. Conventional models, such as DeepLab, rely on detailed pixel-level labels that take time and effort from medical professionals to generate. These labels are sometimes difficult to obtain, particularly during urgent health crises. This research applies a weakly supervised method that uses text descriptions of lesions from X-ray images to generate effective segmentation without pixel-level labels. The proposed method incorporates a classification framework that differentiates between positive and negative text expressions to localize target regions within the chest X-ray images. Text from the same image is treated as positive. In contrast, text from unrelated images acts as negative input, allowing the model to focus on relevant features without relying on explicit pixel-wise labeling. Using this method, labels can be generated more efficiently by leveraging text descriptions of medical images. This approach significantly reduces the workload on healthcare providers and offers a more efficient solution for COVID-19 detection. This study demonstrates the potential of weakly-supervised learning in medical image analysis scenarios where labeled data is scarce. By leveraging deep learning techniques, this research contributes to improving COVID-19 diagnosis efficiency. It highlights the potential of AI in addressing urgent healthcare challenges, potentially speeding up the diagnostic process in critical situations.