Lesion Segmentation in Gastroscopic Images Using Generative Adversarial Networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of gastric lesions in gastroscopic images is crucial for early gastric cancer detection and treatment.
• Existing segmentation models may lack the precision required for clinical applications.

Purpose of the Study:

• To propose a novel generative adversarial training approach for enhanced gastric lesion segmentation.
• To improve the accuracy and reliability of automated segmentation in gastroscopic images.

Main Methods:

• A U-Net based segmentation network incorporating residual blocks and cascaded dilated convolutions was developed.
• A Markov discriminator (Patch-GAN) was employed to differentiate generated from real segmentation masks.
• Adversarial training iteratively optimized the generator and discriminator for improved performance.

Main Results:

• The proposed method achieved a Dice score of 86.6%, accuracy of 91.9%, and recall of 87.3%.
• These results demonstrate significant improvements over existing models.
• The enhanced segmentation performance indicates clinical utility for diagnosis and treatment planning.

Conclusions:

• The novel generative adversarial approach effectively segments gastric lesions in endoscopic images.
• The method shows promise for improving early gastric cancer diagnosis and treatment.
• The integration of residual connections and dilated convolutions enhances context integration and information propagation.