AI-Assisted Detection of Early Gastric Cancer via Visualization of Mucosal Acidity Compromise During Endoscopy

Accurate localization of early gastric cancer (EGC) remains challenging due to its morphological resemblance to gastritis. This study presents an artificial intelligence (AI)-assisted bedside diagnostic system to enhance EGC detection by visualizing gastric mucosal acidity. The ATPase H⁺/K⁺ transport β subunit (ATP4B), a key regulator of acid secretion, is progressively downregulated in gastric mucosal atrophy and intestinal metaplasia, and significantly reduced in EGC. A surface-enhanced Raman scattering (SERS) microarray is developed to map mucosal pH in 50 patient specimens (1,516 points), with founding compared to pathological images. A multi-model neural network is trained and validated internally on data from 40 patients (1,127 points) and externally validated on 10 patients (389 points). Using an optimal pH threshold of 6.845, the system achieved a strong correlation (R² = 0.79) and low error (SSE = 71.83). External validation demonstrated 87.79% sensitivity, 85.04% specificity, 86.89% accuracy, and a κ score of 0.71. This system detected mild pH shifts in atrophic gastritis with intestinal metaplasia, but marked increases with EGC onset, and is able to predict inflammation prior to pathology confirmation. By integrating pH mapping with morphological features, this approach enables precise EGC localization, improves guidance for endoscopic submucosal dissection (ESD), and reduces false-positive diagnoses.