Impact of network architecture and training strategy for photon counting CT data correction

In recent years, x-ray photon-counting detectors (PCDs) have become increasingly popular due to their ability to discriminate energy and low noise levels. However, technical issues (e.g., charge splitting and pulse pileup effects) can affect the data quality by distorting the energy spectrum. To address those issues, based on a deep neural network-based approach using a Wasserstein generative adversarial network (WGAN) framework for PCD data correction, we evaluate the effectiveness of pre-trained and training-from-scratch convolutional neural networks (CNNs) as perceptual loss functions to address charge splitting and pulse pileup correction challenges in photon counting computed tomography (CT) data. Different CNN architectures, including VGG11, VGG13, VGG16, VGG19, ResNet50, and Xception, are evaluated. Compared with the method using a pre-trained network, our findings indicate that training the CNNs from scratch on our dataset produces better results. It significantly affects the performance for the choice of CNN architecture as a perceptual loss in the WGAN framework. Furthermore, because recent explosive interest on transformers has suggested their potential to be useful for computer vision tasks, we also evaluate transformers to maximize the attribute-related information contained in the image feature by texture features extraction. Our study emphasizes the importance of selecting appropriate network architecture and training strategy when implementing the WGAN framework for photon counting CT data correction.