A Scale-Aware Framework for Simultaneous Segmentation of Pancreas and Tumors

Simultaneous segmentation of the pancreas and tumors is crucial in clinical applications. However, this task remains challenging, especially for pancreatic tumors, due to their extensive phenotypic differences, such as scale variations. To address these challenges, we propose a novel framework to enhance scale awareness at both macro- and micro-levels. At the macro level, we promote the perception of different scale representations by integrating multi-scale patch information via a patch-adjust fusion approach. At the micro-level, we introduce a novel Scale-Aware Network (SA-Net) to efficiently learn scale-aware features through two key scale-aware components in the encoder and decoder, respectively. Specifically, the Hierarchical Context Aggregator (HCA) enhances the robustness of the model to scale variations by using channel attention to adaptively fuse features of different scales in the encoder. The Multi-Scale Attention Mechanism (MSAM) improves the network’s scale-discriminative capabilities by promoting the connections between different decoder features via attention and atrous spatial pyramid pooling operations. Both the HCA (0.19M) and the MSAM (0.06M) are lightweight, facilitating the deployment of our approach in clinical settings. Experimental results on the public MSD dataset demonstrate the superior performance of our proposed segmentation framework compared to other state-of-the-art methods. Ablation studies further confirm the efficacy of the proposed scale-aware modules, underscoring their critical role in enhancing segmentation accuracy.