Accurate object localization facilitates automatic esophagus segmentation in deep learning.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Automatic esophagus segmentation is difficult due to the organ's small size, low contrast, and variable shape.
• Existing deep learning methods face challenges in accurately segmenting the esophagus.

Purpose of the Study:

• To enhance esophagus segmentation performance using a two-stage deep learning approach.
• To improve the robustness and accuracy of automatic esophagus delineation.

Main Methods:

• A modified CenterNet model was used for initial esophagus center localization on thoracic CT scans.
• 3D U-Net and 2D U-Net models were trained for segmentation based on the localized center.
• A fine-tuning step with a 2D U-Net refined segmentation using an updated object center.

Main Results:

• The 2D U-net_fine model achieved the highest mean Dice coefficient (0.82) and a low 95% Hausdorff distance (3.76).
• The two-stage strategy significantly improved segmentation by 5.5% in cases with initial poor results (Dice < 0.75).
• Segmentation performance varied across esophageal regions, with lower accuracy between the inferior orifice and pulmonary bifurcation.

Conclusions:

• A two-stage strategy combining object localization and segmentation enhances deep learning model robustness.
• Accurate initial object localization is crucial for significantly improving esophageal delineation, especially in difficult cases.
• The proposed method offers a promising solution for challenging automatic esophagus segmentation tasks.