CT Image Conversion among Different Reconstruction Kernels without a Sinogram by Using a Convolutional Neural Network

This study aimed to develop and validate a convolutional neural network (CNN) to convert CT images reconstructed with one kernel to those with different reconstruction kernels without using a sinogram. The study included ten chest CT scans reconstructed with B10f, B30f, B50f, and B70f kernels, divided into training, validation, and testing datasets. A CNN with six convolutional layers was constructed. Performance was evaluated using root mean square error (RMSE) values. For clinical validation, 30 additional chest CT scans reconstructed with B30f and B50f kernels were converted and analyzed using Aview software for emphysema quantification. The scheme achieved a rapid conversion rate of 0.065 s/slice and significantly reduced RMSE (mean reduction of 65.7%). Emphysema indices for B30f, B50f, converted B30f, and converted B50f were 5.4%, 15.3%, 5.9%, and 16.8%, respectively. The CNN-based conversion demonstrated high accuracy and speed, highlighting its potential for clinical use.