Predicting ventilation from single breathing phase non-contrast CT using Swin Transformers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Pulmonary ventilation imaging quantifies lung function for radiotherapy planning.
• Current SPECT imaging uses radioactive tracers, causing artifacts.
• CT ventilation imaging (CTVI) typically requires multiple CT phases (e.g., 4DCT).

Purpose of the Study:

• To develop a deep learning model predicting SPECT ventilation from single-phase inhale CT.
• To evaluate the benefit of including exhale CT phase data.
• To simplify CTVI for broader clinical feasibility and reduced computational cost.

Main Methods:

• Developed a SwinUNETR-based deep learning model for ventilation prediction.
• Trained and compared models using inhale CT, exhale CT, and inhale/exhale CT (IECT) pairs.
• Evaluated performance against standard U-Net architectures.

Main Results:

• The hybrid SwinUNETR model (IECT-Swin-FTD) achieved state-of-the-art accuracy (Spearman correlation 0.762).
• Single-phase inhale CT model showed comparable results to multi-phase U-Net models.
• Transformer-based models outperformed U-Net architectures.

Conclusions:

• Transformer-based CTVI models can achieve state-of-the-art SPECT ventilation prediction.
• Single-phase CT is potentially sufficient, improving clinical workflow efficiency.
• Reduced acquisition and registration artifacts enhance CTVI feasibility.