TriSwinUNETR lobe segmentation model for computing DIR-free CT-ventilation.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Functional radiotherapy aims to spare high-ventilation lung regions from high radiation doses.
• Current CT-ventilation imaging (CTVI) methods often rely on deformable image registration (DIR), which can introduce inaccuracies.
• Accurate lung lobe segmentation is crucial for precise CTVI and radiotherapy planning.

Purpose of the Study:

• To develop and validate an AI-driven method for calculating CTVI using lung lobe segmentations, eliminating the need for DIR.
• To hypothesize that AI-driven lobe segmentation can accurately determine lung volume changes for CTVI.
• To utilize a novel TriSwinUNETR pipeline for lobe segmentation and subsequent CTVI calculation.

Main Methods:

• A novel lobe segmentation pipeline, TriSwinUNETR, employing three SwinUNETR networks trained on extensive CT image data (COPDGene study).
• The pipeline segments individual lung lobes from inhale/exhale CT pairs to calculate breathing-induced volume changes.
• Validation involved comparing segmentation accuracy using LUNA16 dataset and assessing CTVI correlation with PET-Galligas ventilation in lung cancer patients.

Main Results:

• TriSwinUNETR achieved a state-of-the-art mean Dice score of 93.72% for lobe segmentation, surpassing previous benchmarks.
• CTVI calculations derived from AI-segmented lobes showed a median Spearman correlation coefficient of 0.9 with PET-Galligas ventilation.
• 13 out of 19 patients demonstrated strong agreement (correlation ≥ 0.5) between AI-based CTVI and PET-Galligas ventilation.

Conclusions:

• The TriSwinUNETR pipeline offers superior lobe segmentation accuracy for CTVI.
• Segmentation-based CTVI demonstrates strong agreement with established PET-Galligas ventilation imaging.
• This AI-driven approach reduces reliance on DIR, providing interpretable results and facilitating quality assurance in radiotherapy planning.