Accurate segmentation of lung lesions from the CT scans is crucial for assessing disease severity and tracking the treatment response. However, manual delineation is slow, subjective, and impractical in emergency clinical settings. Automated segmentation is therefore essential for rapid and consistent analysis. Recently, U-Net and its variants have significantly improved the medical image segmentation, but challenges such as low lesion contrast, irregular boundaries, and class imbalance limit their performance on clinical data. To address these issues, this study builds upon the U-Net framework, introducing two lightweight variants: Progressively Regularized Attention U-Net (PRA-Net) and Hierarchical Pervasive Attention Network (HPA-Net). These models are designed to enhance lesion boundary precision and generalization while remaining computationally efficient. Using the publicly available COVID-19 CT dataset, pre-processing techniques including CLAHE (Contrast Limited Adaptive Histogram Equalization) based contrast enhancement, Gaussian smoothing, and augmentation were applied to improve the lesion visibility. PRA-Net employs progressive dropout regularization, while HPA-Net integrates attention layers at multiple depths along with a self-attention bottleneck for global feature refinement. PRA-Net achieved the highest Dice (0.86) and IoU (0.75) score, while HPA-Net recorded the lowest validation loss (128.4). These results demonstrate the potential of attention driven architectures for fast, reliable CT-based lesion quantification in clinical environments.