CT ventilation images produced by a 3D neural network show improvement over the Jacobian and HU DIR-based methods to...

Area of Science:

• Medical Imaging and Radiation Oncology
• Artificial Intelligence in Healthcare
• Computational Biology and Bioinformatics

Background:

• Radiation-induced pneumonitis is a significant risk for non-small cell lung cancer (NSCLC) patients, with incidence up to 33% and mortality at 2%.
• Current radiotherapy planning assumes homogeneous lung function, potentially irradiating critical high-functioning lung regions.
• Functional lung avoidance, by sparing high-functioning lung areas, shows promise in reducing pneumonitis risk.

Purpose of the Study:

• To develop a more accurate method for creating CT ventilation images (CTVI) to identify functional lung regions for radiotherapy planning.
• To overcome limitations of existing CTVI methods, such as deformable image registration (DIR) inaccuracies and sensitivity to artifacts.
• To utilize a neural network approach for predicting ventilation maps directly from breath-hold CT (BHCT) images.

Main Methods:

• A 3D neural network (nnU-Net) was trained using a five-fold cross-validation ensemble to predict ventilation maps (CTVI_nnU-Net) from BHCT images.
• Training data included registered BHCT and Galligas PET images from 20 patients, with ground truth derived from quantized PET intensities (high, medium, low function).
• Performance was compared against a 2D U-Net (CTVI_nnU-Net-2D) and DIR-based methods (CTVI_Jac, CTVI_HU), evaluating Dice Similarity Coefficient (DSC) and Hausdorff Distance 95th percentile (HD95).

Main Results:

• The 3D neural network (CTVI_nnU-Net) demonstrated superior similarity to the ground truth PET data, achieving a mean DSC of 0.68.
• CTVI_nnU-Net showed comparable or lower spatial distance to functional regions (HD95 of 22 mm) compared to other methods.
• The 3D neural network outperformed 2D U-Net and DIR-based methods in both similarity (DSC) and spatial accuracy (HD95) metrics.

Conclusions:

• The developed 3D neural network accurately generates quantized CTVI, outperforming existing 2D U-Net and DIR-based techniques.
• Direct CTVI prediction by the neural network eliminates the need for post-processing thresholding to identify high-function lung areas.
• The neural network approach offers improved accuracy and faster evaluation, indicating significant potential for clinical implementation in functional lung avoidance strategies.