Synthetic pulmonary perfusion images from 4DCT for functional avoidance using deep learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiotherapy Planning

Background:

• Accurate pulmonary perfusion imaging is crucial for radiotherapy planning in lung cancer.
• Current methods like SPECT/CT are resource-intensive and involve radiation exposure.
• Synthetic imaging offers a potential non-invasive alternative.

Purpose of the Study:

• To develop and evaluate a deep learning model for generating synthetic pulmonary perfusion images.
• To utilize 4DCT (four-dimensional computed tomography) scans as input for the model.
• To assess the performance of synthetic images compared to actual perfusion studies.

Main Methods:

• A 3D-residual network was trained on a dataset of 58 99mTc-MAA-SPECT/CT perfusion studies and 4DCT scans.
• The model used inhale and exhale phases of 4DCT to predict perfusion, with MAA-SPECT serving as ground truth.
• Performance was evaluated using correlation coefficients and contour agreement metrics on a hold-out test set.

Main Results:

• The deep learning model achieved a Spearman correlation of 0.70 and Pearson correlation of 0.66 for perfusion prediction.
• Agreement for functional avoidance contours showed a Dice score of 0.803 and average surface distance of 5.92 mm.
• The model demonstrated robust performance on the test set.

Conclusions:

• Deep learning can accurately generate synthetic pulmonary perfusion images solely from 4DCT data.
• This technology holds potential for improving functional avoidance in radiotherapy treatment planning.
• Synthetic perfusion imaging may reduce the need for additional SPECT/CT scans.