Real-time lung extraction from synthesized x-rays improves pulmonary image-guided radiotherapy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate localization of lung tumors in X-rays is challenging due to obscured visibility.
• Overlapping thoracic structures like ribs and the heart hinder precise identification of lung abnormalities.

Purpose of the Study:

• To develop and validate a lung extraction network (LeX-net) for enhancing lung conspicuity in anterior-posterior (AP) X-rays.
• To improve the accuracy of lung tumor tracking in image-guided procedures by virtually removing overlapping thoracic structures.

Main Methods:

• Trained a Swin-UNet model with conditional GAN (LeX-net) using digitally reconstructed radiographs (DRRs) from 73,965 3DCTs and 106 4DCTs.
• Input: full anatomy DRRs; Reference: lung-extracted DRRs. LeX-net generated lung-extracted DRRs from full anatomy DRRs.
• Evaluated LeX-net performance on 4DCT-derived DRRs and assessed tumor tracking accuracy using a template-matching method on a holdout 4DCT test set.

Main Results:

• LeX-net achieved high performance in isolating lungs in DRRs (SSIM: 0.9581 ± 0.0151, PSNR: 30.78 ± 2.50) on 3DCT data.
• When applied to 4DCT data, performance slightly decreased but maintained usable lung-only views.
• Significantly reduced tumor tracking error (0.97 ± 0.86 mm vs. 3.13 ± 5.82 mm) and improved tracking success rates across different window sizes.

Conclusions:

• LeX-net effectively removes overlapping anatomies, enhancing lung visualization in X-rays.
• The model generalizes from 3DCT to 4DCT-derived DRRs, demonstrating robustness.
• LeX-net significantly improves lung tumor tracking outcomes, aiding image-guided interventions.