Abstract : The computational complexity and memory requirements of architectures like UNet make it difficult to deploy deep learning models for medical picture segmentation efficiently. In this work, we propose a lightweight structured pruning approach that leverages the scaling factor (γ) from Batch Normalization layers as a proxy for activation strength to guide filter importance. The traditional pruning techniques rely on activation maps or gradients where expensive forward passes are required which can increase memory overhead. Our method enables pruning during or after training without external profiling. Inspired by Network Slimming defined by Liu et al. (2017), we integrate γ-based ranking with L1-norm, precomputed FLOPs, and sensitivity metrics into a unified Dynamic Weighted Multi-Objective Scoring (DW-MOS-lite) framework. Using a custom UNet trained on lung CT scans, our approach maintains a high Dice coefficient of 90.49% while pruning almost 50% of parameters. It is therefore perfect for edge-based, real-time medical applications. The results demonstrate that Batch Norm γ is a highly efficient and interpretable activation proxy for structured pruning.