Development and validation of a novel YOLOv5-based artificial intelligence model for gastric mucosal lesion detection.

BACKGROUND AND AIMS: Artificial intelligence (AI) has been widely used in endoscopic diagnosis; however, an AI model capable of comprehensively diagnosing both diffuse and focal lesions remains lacking. This study aimed to develop an AI-based endoscopic analysis system, Endosmart, using the YOLOv5 framework to enable synchronous detection of diffuse and focal gastric lesions.

METHODS

A total of 34,979 gastroscopic images were collected from two tertiary hospitals. The diagnostic performance of Endosmart was assessed using accuracy, sensitivity, specificity, and F1-score. The diagnostic performance of the model was assessed by the ROC curve and the AUC. External validation was conducted on 953 gastroscopic images independently assessed by Endosmart and four competent endoscopists. Moreover, the consistency between Endosmart and the average performance of the competent endoscopists in lesion classification was evaluated.

RESULTS

Endosmart demonstrated robust performance in both independent and external validation sets. In the independent validation, the model achieved an overall AUC of 0.965, with AUC exceeding 0.950 across all mucosal lesion types. In the external validation, Endosmart maintained high diagnostic accuracy across lesion types. For focal low-risk lesions, accuracy was 0.960 (0.948-0.973), sensitivity 0.957 (0.941-0.972), and specificity 0.966 (0.944-0.986). For focal high-risk lesions, accuracy was 0.965 (0.954-0.977), sensitivity 0.924 (0.876-0.967), and specificity 0.972 (0.961-0.983). For diffuse lesions, accuracy reached 0.990 (0.982-0.995), with sensitivity of 0.987 (0.966-1.000) and specificity of 0.990 (0.983-0.996). Notably, Endosmart demonstrated robust diagnostic performance compared to competent endoscopists across different lesion types (p < 0.05). Specifically, for focal low-risk lesions, Endosmart showed better performance across all evaluated metrics. For focal high-risk lesions, Endosmart achieved higher accuracy, sensitivity, and NPV, with comparable specificity and PPV. For diffuse lesions, Endosmart achieved higher accuracy, whereas endoscopists showed superior specificity and PPV.

CONCLUSIONS

Endosmart enables efficient and simultaneous diagnosis of diffuse and focal lesions in upper gastrointestinal endoscopy.