Perfusion estimation from dynamic non-contrast computed tomography using self-supervised learning and a...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Diagnostics

Background:

• Pulmonary perfusion imaging is crucial for lung health assessment but limited by current nuclear medicine techniques.
• Existing methods suffer from low spatial resolution and long acquisition times, restricting clinical use and increasing costs.

Purpose of the Study:

• To develop a novel deep learning approach for predicting pulmonary perfusion imaging.
• To utilize non-contrast inhale and exhale computed tomography (IE-CT) scans as input for the prediction model.

Main Methods:

• A U-Net Transformer architecture was developed, modified for Siamese IE-CT inputs.
• Self-supervised learning on 523 IE-CT images was used to learn a low-dimensional feature space.
• Supervised training with transfer learning was performed on 44 patients with IE-CT and SPECT/CT perfusion scans.

Main Results:

• The deep learning model achieved a state-of-the-art spatial Spearman correlation of 0.742 ± 0.037 with ground truth SPECT perfusion.
• A mean median correlation of 0.792 ± 0.036 was observed, indicating high prediction accuracy.

Conclusions:

• The novel approach effectively combines inhale and exhale CT features using deep learning, aligning with physical modeling principles.
• This method shows potential for faster, more accurate lung function imaging, expanding clinical applications beyond nuclear medicine.