Advancing Barrett's Esophagus Segmentation: A Deep-Learning Ensemble Approach with Data Augmentation and Model Collaboration.

ABSTRACT

This approach provides a thorough investigation of Barrett's esophagus segmentation using deep-learning methods. This study explores various U-Net model variants with different backbone architectures, focusing on how the choice of backbone influences segmentation accuracy. By employing rigorous data augmentation techniques and ensemble strategies, the goal is to achieve precise and robust segmentation results. Key findings include the superiority of DenseNet backbones, the importance of tailored data augmentation, and the adaptability of training U-Net models from scratch. Ensemble methods are shown to enhance segmentation accuracy, and a grid search is used to fine-tune ensemble weights. A comprehensive comparison with the popular Deeplabv3+ architecture emphasizes the role of dataset characteristics. Insights into training saturation help optimize resource utilization, and efficient ensembles consistently achieve high mean intersection over union (IoU) scores, approaching 0.94. This research marks a significant advancement in Barrett's esophagus segmentation.