Generalizable cone beam CT esophagus segmentation using physics-based data augmentation.

Area of Science:

• Medical Physics
• Radiotherapy
• Medical Imaging

Background:

• Accurate esophagus segmentation is vital in lung cancer radiotherapy to mitigate radiation-induced esophagitis.
• Current segmentation methods face challenges with image quality variations between planning CT (pCT) and cone beam CT (CBCT).

Purpose of the Study:

• To develop and validate a semantic physics-based data augmentation technique for robust esophagus segmentation.
• To improve segmentation accuracy across different imaging modalities (pCT and CBCT) in adaptive radiotherapy.

Main Methods:

• A modified 3D U-Net architecture and multi-objective loss function were employed.
• Physics-based artifact induction was used to generate synthetic CBCT data from pCT scans.
• The method incorporated scatter artifacts and noise characteristic of CBCT.

Main Results:

• The model achieved state-of-the-art Dice overlaps of 0.81 on pCT and 0.74 on CBCT.
• The physics-based augmentation demonstrated robustness and generalizability across datasets and modalities.
• The approach effectively addressed noise and artifacts present in CBCT images.

Conclusions:

• Physics-based data augmentation is effective for segmenting soft-tissue organs like the esophagus.
• The method generalizes well across imaging modalities, enhancing treatment setup accuracy.
• This technique has the potential to improve response analysis in radiotherapy.