Development and Validation of an Explainable Artificial Intelligence Model for Early Gastric Cancer Diagnosis Using Multimodal Endoscopic Imaging
Development and Validation of an Explainable Artificial Intelligence Model for Early Gastric Cancer Diagnosis Using Multimodal Endoscopic Imaging
Early gastric cancer (EGC) is often difficult to detect accurately during endoscopic examination due to subtle morphological features and variability among endoscopists. Artificial intelligence (AI) has shown promise in improving diagnostic performance; however, most existing models lack interpretability and rely on single-modality imaging.
This study aims to develop and evaluate an explainable multimodal artificial intelligence model for the diagnosis of early gastric cancer using endoscopic imaging. The model integrates features derived from white-light imaging and image-enhanced endoscopy, along with quantitative image features and clinical data, to improve diagnostic accuracy and provide interpretable decision support.
The primary outcome is the diagnostic performance of the AI model for detecting early gastric cancer, evaluated by area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity.
The results of this study are expected to provide evidence for the clinical utility of explainable AI in endoscopic diagnosis and support the development of reliable human-AI collaborative diagnostic systems.
This study is designed to develop and evaluate an explainable multimodal artificial intelligence (AI) model for the diagnosis of early gastric cancer (EGC) based on endoscopic imaging.
Patients who underwent endoscopic submucosal dissection (ESD) will be collected from one or more medical centers. Eligible patients will include those with histopathologically confirmed early gastric cancer or non-cancerous lesions who have undergone both white-light imaging (WLI) and image-enhanced endoscopy. For each lesion, representative endoscopic images will be selected according to predefined quality criteria.
The study will be conducted in several steps. First, a lesion detection model will be developed to identify regions of interest in endoscopic images. Second, quantitative image features, including color, texture, morphological, and mucosal structural features, will be extracted from the detected regions. Third, deep learning models will be used to derive predictive features from both WLI and image-enhanced endoscopic images. These features will be integrated with clinical variables to construct a multimodal prediction model.
The dataset will be divided into training, validation, and testing subsets. Model performance will be evaluated using standard diagnostic metrics, including the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, and F1 score. Calibration performance and clinical utility may also be assessed using calibration curves and decision curve analysis.
To enhance model interpretability, feature contribution will be analyzed using Shapley additive explanations (SHAP), and visualization techniques such as gradient-weighted class activation mapping (Grad-CAM) will be applied to highlight regions of interest contributing to model predictions.
This study aims to develop an accurate and interpretable AI-based diagnostic tool for early gastric cancer and to provide evidence supporting its potential application in clinical endoscopic practice.
Inclusion Criteria:
Exclusion Criteria:
lhliu2024@stu.suda.edu.cn+8615943593759