C72-03 Convolution-Based Computer Tomography Ventilation

Abstract Rationale Computer Tomography Ventilation (CTV) algorithms use four-dimensional CT (4DCT) data to generate three-dimensional maps of lung volume changes between an inhale/exhale image pair. Clinicians use these ventilation maps to visualize regional lung function. Intensity-based CTV algorithms estimate material density and corresponding volume change from differences in CT Hounsfield Units (HU). Early methods compared HU values at spatially corresponding inhale and exhale voxels identified by deformable image registration (DIR), but DIR variability significantly reduced CTV reproducibility. Later methods improved reliability by averaging intensities over adaptive subregions and solving non-trivial optimization problems to recover voxel-wise volume change. However, the complexity of such methods has limited clinical adaptation. We introduce a novel convolution-based CTV algorithm that is robust to DIR variability, while remaining simplistic and computationally light compared to existing approaches. Methods Our proposed method replaces adaptive subregions with a uniform convolution kernel (K), applying three-dimensional convolutions to each CT image pair and associated binary lung masks. We use DIR to map inhale/exhale image pairs, then apply lung masks to isolate parenchyma. To exclude regions outside the expected soft tissue/air range, we convert HU values to density estimates and remove voxels exceeding a density threshold. Each thresholded density image and mask are convolved separately, and their ratio yields the local mean density within K. We repeat this process for inhale and exhale and calculate the voxel-wise CTV image as the ratio of local exhale to inhale densities. Overall, the method requires four straightforward convolution operations. We performed a grid-based parameter search over kernel sizes and threshold values to identify the combination that maximized correlation with ground truth nuclear medicine ventilation scans. Results Fifteen inhale/exhale CT image pairs with corresponding nuclear medicine ventilation scans were used to evaluate the proposed method. Voxel-wise Spearman correlation coefficients were computed between the estimated ventilation and the nuclear medicine scans. Across all cases, the highest average performance was achieved using a threshold value of 0.4, and a voxel kernel size of 21, 7, representing an isotropic box in millimeter space. This yielded a median correlation of 0.606, with the 25th percentile at 0.455 and the 75th at 0.711. The lowest correlation observed was 0.121, and the highest was 0.799. Conclusions Numerical results show that the convolution-based method yields consistent, reproducible ventilation estimates that agree well with nuclear medicine scans, regardless of DIR variability. Findings suggest that this approach offers a practical, efficient alternative to existing HU-based techniques. This abstract is funded by: Funded by NIH/NHLBI R01HL169869