Technical Note: Deriving ventilation imaging from 4DCT by deep convolutional neural network.

Area of Science:

• Medical Imaging
• Computational Biology
• Radiology

Background:

• Four-dimensional computed tomography (4DCT) enables ventilation imaging by analyzing changes in CT values across respiratory phases.
• Deformable image registration (DIR) is crucial for accurate 4DCT-derived ventilation images but introduces algorithm-dependent uncertainties.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) method for deriving ventilation images directly from 4DCT data.
• To eliminate the need for explicit deformable image registration (DIR) in 4DCT-based ventilation imaging.
• To reduce the uncertainty associated with DIR in the generation of ventilation images.

Main Methods:

• A deep CNN architecture was designed using 82 sets of 4DCT and ventilation images from lung cancer patients.
• The CNN input comprised two-channel CT data from end-exhale and end-inhale phases.
• Mean-squared error (MSE) was employed as the loss function to compare predicted and reference ventilation images.

Main Results:

• The CNN-generated ventilation images demonstrated high comparability with reference label images.
• Quantitative evaluation showed a similarity index of 0.880 ± 0.035, a correlation coefficient of 0.874 ± 0.024, and a Gamma index passing rate of 0.806 ± 0.014.
• These metrics were averaged over tenfold cross-validation, indicating robust performance.

Conclusions:

• Deep CNNs can effectively generate ventilation images from 4DCT without requiring explicit deformable image registration.
• The proposed CNN-based approach successfully mitigates the uncertainties inherent in traditional DIR methods for ventilation imaging.