Parallel deep neural networks for endoscopic OCT image segmentation.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of endoscopic optical coherence tomography (OCT) images is crucial for diagnosing esophageal conditions.
• Limited training datasets pose a challenge for developing robust deep learning models for OCT image analysis.

Purpose of the Study:

• To develop and validate parallel-trained deep neural networks for automated endoscopic OCT image segmentation.
• To assess the method's performance in layer segmentation and its potential for clinical applications, such as differentiating disease states.

Main Methods:

• U-Net-based deep neural networks were trained using a modified dice loss function and manual segmentations.
• Ultrahigh-resolution cross-sectional images were acquired using an 800 nm OCT endoscopic system.
• The method was tested on *in vivo* guinea pig esophagus images and applied to an eosinophilic esophagitis (EOE) model.

Main Results:

• The deep neural networks achieved robust layer segmentation with a low boundary error of 1.4 µm, independent of topological variations.
• The method successfully differentiated *in vivo* OCT esophagus images from an EOE model and control group.
• Quantitative changes in esophageal layer thickness were clearly demonstrated in the EOE model.

Conclusions:

• Parallel-trained deep neural networks offer a feasible solution for automated endoscopic OCT image segmentation, even with limited data.
• The developed method demonstrates high accuracy and robustness for esophageal layer segmentation.
• This approach shows significant clinical potential for diagnosing and monitoring esophageal diseases like EOE.