Medical image segmentation remains a critical challenge, particularly due to the scarcity of annotated data and stringent patient confidentiality regulations. This paper presents a novel unsupervised learning approach to segment COVID-19 computed tomography (CT) scans without requiring pixel-level annotations. Our method leverages a 3D Generative Adversarial Network (GAN) to synthesize healthy CT images from COVID-19 scans. By subtracting these generated healthy images from the original CT scans, we create pseudo-masks that highlight infected regions. These pseudo-masks are then utilized to train a U-Net segmentation network using a contrastive loss function. Our approach effectively mitigates the need for manual annotation and demonstrates promising segmentation performance. This method offers a viable solution for scenarios with limited annotated data and emphasizes the potential of unsupervised learning techniques in medical imaging.