Gastric polyp detection in gastroscopic images using deep neural network

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastric polyps vary in size, posing detection challenges, especially for smaller ones.
• Accurate polyp detection in gastroscopic images is crucial for early diagnosis and treatment.
• Existing methods struggle with the subtle features of small gastric polyps against complex backgrounds.

Purpose of the Study:

• To develop and evaluate a deep learning object detection method for improved gastric polyp detection.
• To specifically address the challenge of detecting small gastric polyps in endoscopic images.
• To enhance the fusion of multi-level features for better polyp identification.

Main Methods:

• A novel feature extraction and fusion module was developed.
• The module was integrated with the YOLOv3 (You Only Look Once version 3) object detection network.
• The model was trained and validated on a custom dataset of 1433 training and 508 validation gastroscopic images.

Main Results:

• The proposed method demonstrated superior performance in detecting small gastric polyps compared to other approaches.
• The feature fusion module effectively combined high-level semantic and low-level detailed information.
• Achieved high performance metrics: 91.6% precision, 86.2% recall, 88.8% F1 score, and 87.2% F2 score.

Conclusions:

• The integrated deep learning model significantly improves gastric polyp detection accuracy, particularly for small polyps.
• The feature extraction and fusion module is key to enhancing the identification of subtle polyp features.
• This advanced method offers a promising tool for augmenting gastrointestinal diagnostics.