Abstract
Attention-augmented encoder-decoder models with non-local modules are effective for lung tumor analysis on computed tomography (CT), yet their computational and memory demands can hinder deployment. We study pruning for a specialized non-local U-Net pipeline that jointly performs neoplasm presence recognition and lesion segmentation from single-channel CT snapshots. We benchmark unstructured and structured pruning baselines and propose a sensitivity-aware, architecture-guided refinement that allocates sparsity across encoder-decoder components and attention modules to better preserve task-critical capacity. On a joint benchmark of public lung CT datasets, the proposed method achieves a superior accuracy-efficiency trade-off. At 50% sparsity, it reduces latency from 41.5 ms to 23.6 ms and FLOPs from 62.4 G to 30.8 G while maintaining strong quality$(\mathrm{F} 1=0.923, \text{mIoU}=0.742,\ \text{Dice}=0.847$). At 60 % sparsity, it further reduces latency to$19.4 \mathrm{ms} (\text{FLOPs} = 23.5 \mathrm{G})$with stable performance$(\mathrm{F} 1=0.904, \text{mIoU}=0.724)$. Overall, the proposed pruning refinement improves the deployability of attention-based lung CT models without sacrificing clinically meaningful performance.