A deep learning model for the diagnosis of gastric neuroendocrine carcinoma
Tianchen Zhu, Zihan Zhao, C. Wang, Xinke Zhang, Lin Zheng, Wenxu Chen, Zhengyi Zhou, Zhiwei Liao, Yan Huang, Muyan Cai, Junpeng Lai · Communications Medicine · 2026
Gastric neuroendocrine carcinoma (G-NEC) presents with clinical and pathological features that closely resemble those of gastric adenocarcinoma (GC), often complicating differential diagnosis. However, G-NEC is markedly more aggressive and associated with a significantly poorer prognosis, necessitating accurate and timely identification to guide appropriate therapeutic interventions. In response to this clinical need, we developed G-NECNet, a deep convolutional neural network tailored to detect G-NEC from histopathological whole-slide images. The model demonstrates excellent diagnostic performance, yielding an average area under the receiver operating curve (AUROC) of 0.993 in the internal validation cohort, 0.985 on an external single-institutional dataset, and 1.000 on an external multi-institutional consultation dataset. These consistently high AUROC values highlight the robustness, accuracy, and generalizability of G-NECNet across diverse clinical settings. The integration of G-NECNet into routine diagnostic workflows may not only improve the precision of G-NEC classification but also reduce misdiagnosis-related healthcare costs, offering a practical and scalable solution for clinical application. Zhu, Zhao, Wang et al. describe G-NECNet, a deep learning model for detecting gastric neuroendocrine carcinoma (G-NEC) from H&E-stained biopsy whole-slide images without the need for additional immunohistochemistry. G-NECNet achieves high diagnostic accuracy across multiple datasets. This study aims to improve the diagnosis of gastric neuroendocrine carcinoma (G-NEC), a rare but aggressive stomach cancer often mistaken for the more common stomach cancer called gastric adenocarcinoma. Since accurate diagnosis typically necessitates time-consuming and resource-intensive staining of sections from the tumor, we developed G-NECNet, a computational model that analyzes routine tumor images to distinguish G-NEC precisely. The model works well across multiple datasets, demonstrating high reliability and generalizability in different clinical settings. These findings suggest that G-NECNet could assist pathologists in making faster and more accurate diagnoses of G-NEC. Integrating this tool into routine diagnostic workflows may help reduce errors, improve patient outcomes, and lower healthcare costs, providing a practical and scalable solution for clinical application.