Scale-View Co-Awareness Framework for Simultaneous Segmentation of Pancreas and Tumors

Quantitative measurement of the pancreas and tumors from medical image segmentation is crucial for disease assessment and surgical planning. Despite advances in deep learning for automatic volume measurement, accurate segmentation of the pancreas and tumors remains challenging due to significant phenotypic variations. To address these issues, we propose a scale-view co-awareness segmentation framework termed SVCA-Net, which simultaneously learns scale-aware and view-aware features to enhance pattern-agnostic representations. Notably, SVCA-Net enhances scale awareness at both macro and micro levels. At the macro level, patch-based and global-based segmentation approaches focus on capturing fine-scale details and contextual information, respectively. At the micro level, we introduce the Scale-Aware Network (SA-Net), which incorporates two key components in the encoder and decoder to learn scale-related features efficiently. Specifically, the Hierarchical Context Aggregator (HCA) aggregates multi-level features in the encoder to improve robustness to scale variations, while the Multi-Scale Aware Module (MSAM) integrates features across different decoder levels to enhance scale discrimination. Both HCA (0.044 M) and MSAM (0.008 M) are lightweight, ensuring efficient deployment in clinical settings. For view awareness, we propose an adaptive view fusion method that dynamically adjusts the training sample flow, facilitating the exchange of information from multiple perspectives and reducing storage requirements. Experimental results on the public MSD dataset demonstrate that SVCA-Net outperforms other state-of-the-art methods, achieving DSC scores of 81.70% for the pancreas and 59.43% for pancreatic tumors. On the BraTS 2020 dataset, SVCA-Net delivers competitive segmentation performance, with DSC scores of 91.22%, 86.72 %, and 80.39% for the whole tumor, tumor core, and enhancing tumor, respectively.