Adversarial convolutional network for esophageal tissue segmentation on OCT images.

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of esophageal optical coherence tomography (OCT) images is crucial for disease diagnosis, providing insights into tissue shape and thickness.
• Current deep convolutional network methods struggle with segmentation accuracy due to limited training data and diverse esophageal layer shapes.

Purpose of the Study:

• To propose a novel adversarial convolutional network (ACN) for improved automatic segmentation of esophageal OCT images.
• To enhance the accuracy and robustness of segmentation for better disease diagnosis.

Main Methods:

• Developed an adversarial convolutional network (ACN) framework comprising a generator and a discriminator with U-Net-like architectures.
• Employed adversarial learning where the discriminator performs both real/fake discrimination and pixel classification.
• Utilized adversarial loss to encode high-order pixel relationships, negating the need for post-processing.

Main Results:

• The ACN demonstrated superior performance compared to existing deep learning frameworks in pixel classification accuracy for esophageal OCT images.
• Achieved improved segmentation results, validating the effectiveness of the adversarial training approach.
• Experiments on guinea pig esophageal OCT images confirmed the ACN's capabilities.

Conclusions:

• The proposed ACN framework offers a significant advancement in automatic esophageal OCT image segmentation.
• The method shows potential for clinical applications, such as detecting eosinophilic esophagitis (EoE).
• Adversarial learning effectively addresses limitations of traditional methods, leading to more accurate and reliable segmentation.