- Aims Endoscopic detection of gastric cancer is operator-dependent, and almost all existing artificial intelligence (AI) systems have been developed in high-incidence East Asian settings. We aimed to develop and preliminarily evaluate an AI model for detection and segmentation of gastric cancer in a Western population, where this pathology is low-incidence, using endoscopic submucosal dissectionAims Endoscopic detection of gastric cancer is operator-dependent, and almost all existing artificial intelligence (AI) systems have been developed in high-incidence East Asian settings. We aimed to develop and preliminarily evaluate an AI model for detection and segmentation of gastric cancer in a Western population, where this pathology is low-incidence, using endoscopic submucosal dissection (ESD) histopathology as reference standard. The model was fine-tuned from a previously established semi-supervised network for Barrett’s neoplasia (Meinikheim et al., Endoscopy 2024). Transfer learning is appropriate because both esophageal and gastric cancers share key endoscopic characteristics and multimodal imaging features across WLI, NBI, and TXI. The feature representations learned from large-scale Barrett’s datasets therefore provide a strong initialization. Primary outcomes were Dice similarity coefficient for tumor segmentation and image-level detection sensitivity.
Methods In this retrospective single-centre study at a tertiary-care hospital in Germany (University Hospital Augsburg), we included 84 patients with histologically confirmed gastric adenocarcinoma (T1a or higher) treated by ESD. From these patients, 827 endoscopic images (Olympus systems; WLI, NBI, TXI, with or without indigo chromoendoscopy) showing visible gastric cancer were extracted. Tumor extent was delineated on still images by experts informed by the corresponding ESD specimens and full pathology reports. All data were de-identified before analysis.
A convolutional neural network segmentation model was initialized with weights from the Barrett AI system, which had been pre-trained on 55,273 endoscopic images from 557 patients with Barrett’s esophagus and related neoplasia, and then supervisedly fine-tuned for gastric cancer segmentation. Data were split patient-wise into five train/validation folds (80/20 per fold) to avoid information leakage between patients, and a separate model was trained for each fold. For segmentation, performance was assessed using the Dice coefficient. For image-level detection, a tumor was considered detected in a given image if the entire predicted lesion region overlapped the ground-truth mask with Dice≥75%; detection performance was summarized as sensitivity at this threshold and as the area under a recall–overlap curve obtained by varying the Dice threshold from 0 to 1.
Results Across the five patient-wise folds, tumor Dice averaged 82.93% with a standard deviation of 1.39% [80.91–84.37%]. Image-level detection sensitivity at Dice≥75% averaged 94.86% with a standard deviation of 3.12% [90.23–97.59%]. The area under the recall–overlap curve averaged 0.912 with a standard deviation of 0.009. These results indicate a consistent and robust performance across all validation images.…
