Instance Segmentation of Burn Areas Using Position-Sensitive Convolutional Networks

Abstract We propose a novel end-to-end convolutional network designed for the instance segmentation of burn regions. Building on the foundational principles of fully convolutional networks (FCNs) for semantic segmentation and instance-level mask generation, the method is tailored to address the unique challenges of burn injury assessment. This approach jointly detects and delineates burn areas while maintaining semantic consistency across the region. By incorporating position-sensitive features, such as inside/outside score maps, the network enables precise segmentation, effectively capturing the boundaries and details of burn regions. The architecture ensures shared convolutional representations for both the semantic and instance segmentation tasks, optimizing performance and computational efficiency. This method is highly integrated and demonstrates superior capability in accurately identifying and segmenting burn regions in medical imaging, offering significant potential for clinical applications.