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

Background: Radiation-induced pneumonitis affects up to 33% of non-small cell lung cancer (NSCLC) patients, with fatal pneumonitis occurring in 2% of patients. Pneumonitis risk is related to the dose and volume of lung irradiated. Clinical radiotherapy plans assume lungs are functionally homogeneous, but evidence suggests that avoidance of high-functioning lung during radiotherapy can reduce the risk of radiation-induced pneumonitis. Radiotherapy avoidance structures can be constructed based on high-function regions indicated in a ventilation map, which can be produced from CT images.

Purpose: Existing methods of deriving such a CT ventilation image (CTVI) require the use of deformable image registration (DIR) of peak-inhale and -exhale CT images, which is susceptible to inaccuracy for small or low-intensity regions, and sensitive to image artefacts. To overcome these problems, we use a neural network to predict a ventilation map from breath-hold CT (BHCT).

Methods: We used the nnU-Net pipeline to train five-fold cross-validated ensemble models to predict a ventilation map (CTVI_nnU-Net). The training data were comprised of registered BHCT and Galligas PET images from 20 patients. Three training sets were created to ensure performance was averaged over different test patients. For each set, images from two randomly selected test patients were set aside, and models were trained on the remaining images. The ground truth was established by quantizing the Galligas PET images, assigning each voxel a label of high-function (>70th percentile of intensity), medium-function (between 30th and 70th percentile), or low-function (<30th percentile). For comparison, we created a CTVI with a 2D U-Net (CTVI_nnU-Net-2D), and with the Jacobian (CTVI_Jac) and Hounsfield Units (CTVI_HU) DIR-based methods which we quantized and labeled in the same way. The Dice similarity coefficient (DSC) and Hausdorff Distance 95th percentile (HD95) of each CTVI with the ground truth were measured separately for each lung function subregion.

Results: CTVI_nnU-Net had the highest similarity to the quantized Galligas PET with a mean (range) DSC over all three categories of lung function at 0.68 (0.56 to 0.82), compared with 0.64 (0.47 to 0.75) for CTVI_nnU-Net-2D, 0.60 (0.38 to 0.73) for CTVI_Jac, and 0.56 (0.30 to 0.75) for CTVI_HU. CTVI_nnU-Net had the equal-lowest spatial distance to the quantized Galligas PET averaged over the three categories, with HD95 of 22 mm (9 to 64 mm), compared with 23 mm (9 to 72 mm) for CTVI_nnU-Net-2D, 22 mm (12 to 63 mm) for CTVI_Jac, and 26 mm (12 to 58 mm) for CTVI_HU.

Conclusion: Our 3D neural network produces a quantized CTVI with higher similarity to the ground truth than the 2D U-Net and DIR-based Jacobian and HU methods. As it produces a quantized CTVI directly, CTVI_nnU-Net avoids the need for thresholding to identify high-function lung regions. With faster evaluation and improved accuracy, neural networks show promise for the clinical implementation of functional lung avoidance.