Pleural effusion and ascites, commonly caused by conditions like heart failure and malignancies, require early detection but face challenges due to limited, privacy-sensitive medical data and shortcomings in synthetic data generation. This paper introduces DAMSA (Denoising Autoencoder with Multi-Scale Attention), which integrates Conditional Layer Normalization and adversarial learning to generate high-quality synthetic data while addressing mode collapse and distribution bias. By leveraging multi-scale feature extraction and Dynamic Conditional Control, DAMSA enhances diversity and accuracy in synthetic outputs. Experiments validate its efficacy in dataset expansion and generalization improvement, demonstrating practical utility. Future work will extend DAMSA to broader domains to assess scalability.