Semantic Segmentation of Gastric Polyps in Endoscopic Images Based on Convolutional Neural Networks and an Integrated...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Gastric polyps require accurate identification and removal due to their link to gastric cancer.
• Convolutional neural networks (CNNs) show promise for analyzing endoscopic images.
• Automated diagnosis and segmentation of gastric polyps using CNNs are underexplored.

Purpose of the Study:

• To pioneer research in gastric polyp segmentation within endoscopic images using CNNs.
• To compare the performance of seven classical semantic segmentation models.
• To propose an integrated evaluation approach for optimal model selection.

Main Methods:

• Constructed and compared seven CNN-based semantic segmentation models (U-Net, UNet++, DeepLabv3, DeepLabv3+, PAN, LinkNet, MA-Net) with ResNet50, MobineNetV2, or EfficientNet-B1 encoders.
• Utilized a collected dataset of endoscopic images.
• Developed an integrated evaluation approach combining subjective and objective criteria for model selection.

Main Results:

• UNet++ with the MobineNet v2 encoder demonstrated superior performance based on the integrated evaluation method.
• The selected model was used to build an automated polyp-segmentation system.
• The study confirmed the high clinical value of semantic segmentation models in gastric polyp diagnosis.

Conclusions:

• CNN-based semantic segmentation is highly valuable for gastric polyp diagnosis and treatment.
• The proposed integrated evaluation approach offers an objective tool for selecting among multiple complex models.
• This research advances endoscopic gastrointestinal disease identification and offers a model selection framework for clinical technologies.