Abstract : Deep convolutional neural networks such as UNet have achieved remarkable performance in medical image segmentation; however, these models suffer from high computational complexity and large parameter counts. Pruning is one of the most widely used model compression techniques to address this limitation. The proposed work created a Dynamic Weighted Multi-Objective Scoring (DW-MOS) framework. It consists of correlation-aware structured pruning for an efficient UNet compression that evaluates filter importance by jointly considering weight magnitude, computational cost (FLOPs), layer sensitivity, and inter-filter redundancy. Filter redundancy is explicitly quantified using correlation-based analysis. It helps to remove highly correlated, less informative filters and preserve diverse and complementary features. Also aggregated DW-MOS score helps to apply structured filter pruning in a layer-consistent manner. It also adjusts batch normalization parameters and subsequent convolutional layers to maintain architectural integrity. The Linear interpolation is applied to ensure stable feature alignment after pruning and the pruned network is fine-tuned to recover potential performance loss. Experimental results demonstrate that the dynamic adaptive strategy achieves effective pruning by yielding a 1.16× compression ratio, a 1.21× inference speedup, and a Dice score of 90.98% to detect infection in the lung region of a COVID-19 patient. Hence, the proposed pruning framework is also maintaining clinical relevance.