Abstract
To reflect real-world pathology practice, we developed an artificial intelligence-based pathological computer-aided detection system, trained on diverse epithelial and non-epithelial tumors for gastric biopsy specimens. A multicenter cohort comprising samples from six institutions was used for training and validated with an independent dataset from a seventh institution. We applied two distinct algorithms and three operational validity levels with optimized parameters to address the complexity of pathological diagnosis, reflecting routine diagnostic practice. Our system enabled the detection of malignant regions at low-magnification observation, aligning with the practical workflow of pathologists. In a reader study involving a limited number of test samples, the use of system assistance was associated with an improvement in diagnostic sensitivity. Further analysis revealed that samples with small and dispersed malignant foci had a higher rate of false-negative diagnoses, underscoring the potential of our system to improve diagnostic sensitivity. This study highlights the promise of integrating our system into real-world practice to aid pathologists in the routine diagnosis of gastric biopsy specimens.