Gastric Anatomical Sites Recognition in Gastroscopic Images Based on Dual-branch Perception and Multi-scale Semantic Aggregation

Accurate recognition of key anatomical sites in gastroscopic images is crucial for systematic screening and region-specific diagnosis of early gastric cancer. However, subtle inter-regional differences and indistinct structural boundaries present in gastroscopic images significantly decrease the clinical performance of existing recognition approaches. To address above challenges, we propose a Gastric Anatomical Sites Recognition in Gastroscopic Images Based on Dual-branch Perception and Multi-scale Semantic Aggregation (GASR) for the identification of five representative gastric regions (the greater and lesser curvatures of the antrum, the incisura angularis, and the greater and lesser curvatures of the corpus). Specifically, we propose a Dual-branch Structural Perception (DBSP) module for leveraging the effective complementarity between the local feature extraction of convolutional neural networks (CNNs) and the global semantic modeling of the Swin Transformer. To further improve contextual feature modeling, we develop a Multi-scale Contextual Sampling Aggregator (MCSA) inspired by the Atrous Spatial Pyramid Pooling (ASPP) to extract features across multiple receptive fields. Additionally, we design a Multi-granular Pooling Aggregator (MGPA) based on the Pyramid Scene Parsing (PSP) mechanism to capture hierarchical spatial semantics and global structural layouts through multi-scale pooling operations. Experimental results on a private, expert-annotated endoscopic image dataset, using five-fold cross-validation demonstrate that GASR achieves a recognition accuracy of 97.15%, with robust boundary discrimination and strong generalization performance and can be accepted in clinical practice, showing potential for clinical deployment in gastroscopy-assisted diagnosis and automated screening of early gastric cancer.