DuDeM: A Dual-Network Model for Early Gastric Cancer Detection Based on Capsule Endoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Early detection of gastric cancer significantly improves patient outcomes.
• Capsule endoscopy faces challenges in lesion recognition due to anatomical variations, patient posture, and gastric peristalsis.
• Robust deep learning models are needed to enhance diagnostic accuracy in capsule endoscopy.

Purpose of the Study:

• To develop a deep learning model for reliable early gastric cancer (EGC) detection in capsule endoscopy.
• To address challenges in lesion recognition caused by physiological and anatomical interferences.

Main Methods:

• A dual-network model, DuDeM (DualNet Detection Model), was developed, integrating ResNet50 and CapsuleNet with dynamic routing.
• A convolutional branch extracts local features, feeding into primary capsules for adaptive feature association via dynamic routing.
• Attention-weighted strategy was used for feature fusion, and the model was trained on diverse capsule endoscopy datasets.

Main Results:

• DuDeM achieved a high Area Under the Curve (AUC) of 0.981 and an F1-score of 0.979.
• Sensitivity, specificity, and precision all exceeded 97%, demonstrating strong diagnostic performance.
• The model showed robust performance with less than 3% degradation under mild image perturbations.

Conclusions:

• The DuDeM model enables reliable recognition of early gastric cancer (EGC) in capsule endoscopy.
• Its robustness to image variations suggests potential for large-scale clinical screening applications.
• This deep learning approach can enhance the effectiveness of capsule endoscopy for gastric cancer diagnosis.