Deep learning-aided optical biopsy achieves whole-chain diagnosis of Correa cascade of gastric cancer: a prospective study

BACKGROUND: Biopsies are essential in differentiating benign from malignant lesions in routine gastroscopy. Nevertheless, redundant biopsies increase patients' expenses and pathologists' workload. Probe-based confocal laser endomicroscopy (pCLE) enables real-time in vivo histological evaluation for gastric neoplasms and precancerous conditions. However, endoscopists vary widely in skill, and the use of pCLE requires histopathology expertise, which limits its application in nonacademic settings. This study aimed to develop a pCLE computer-aided diagnosis system (CCADS) for real-time whole-chain diagnosis of Correa cascade of gastric cancer and evaluate it in a real clinical setting. METHODS: Gastric pCLE images and videos from 5771 examinations were retrospectively collected. CCADS was constructed using deep learning networks. It was developed using 47,462 pCLE images and 461 video segments and evaluated via multistep validation. A total of 11,439 images and 667 videos were identified for offline validation. Consecutive patients from October 2019 to September 2021 were enrolled in a prospective diagnostic study for real-time validation, which included 951 patients in the statistics. Blinded expert endoscopists and CCADS independently performed real-time pCLE diagnosis of gastric mucosal lesions in routine examinations, with double-read histopathology as the gold standard. The real-time diagnostic performance of CCADS was evaluated and compared with that of experts. RESULTS: CCADS achieved high diagnostic performance in image test, video test, and a prospective diagnostic study with a large sample size. Overall, 1254 lesions from 951 patients were included in the prospective test. The real-time diagnostic accuracies of CCADS for inflammation, atrophy, gastric intestinal metaplasia (GIM), low-grade intraepithelial neoplasia (LGIN), and high-grade intraepithelial neoplasia and gastric cancer (HGIN/CA) were 91.71%-97.13%. CCADS achieved high sensitivity (98.44%) and specificity (97.06%) for HGIN/CA. Compared with experts, CCADS achieved similar accuracies in diagnosing atrophy, GIM, and LGIN and similar sensitivities in all five categories. Further, CCADS showed a significantly higher sensitivity (96.70% vs. 89.01%, p = 0.044) for gastric neoplasms (LGIN + HGIN/CA) and reduced the misdiagnosis of neoplasms. CONCLUSIONS: CCADS achieved expert-level whole-chain diagnosis of Correa cascade. It could assist endoscopists in improved surveillance of gastric neoplasms and precancerous conditions, promote the application of pCLE, and reduce biopsies. TRIAL REGISTRATION: ClinicalTrials.gov Identifier: NCT03784209.