흉부 CT 영상에서의 재구성 커널 변환을 위한 비지도 적대적 생성 신경망 네트워크 개발

Abstract

Computed tomography (CT) image is one of the diagnostic imaging widely used in the medical field. CT image is reconstructed from sinogram, which is the 2D array data containing the projections, using convolution kernel through back projection. At this point, the kernel differs depending on which anatomical structure is evaluated in qualitative evaluation. Also, quantitative evaluation is crucial as well as qualitative evaluation and affects the choice of kernel. However, there are two problems. First, sinogram has large capacity and storage space is limited, so CT image is usually reconstructed with only one specific kernel for evaluation and sinogram is removed in a week. Second, patients should be scanned and exposed radiation once again. Recently, many researchers have proposed image-to-image translation methods using generative adversarial networks (GANs) for CT kernel conversion. Nevertheless, preserving anatomical structure including fine details, e.g., airway and blood vessel, while transferring the style of the target kernel is still challenging when CT image is translated from the source kernel to the target kernel. In this study, kernel conversion GAN (KCGAN) is proposed to alleviate these problems with perceptual guidance and showed robust and efficient performance in kernel conversion. Perceptual guidance is a type of discriminator regularization method using feature map of generator to learn semantic representation better. For content and style features, cosine similarity content loss and contrastive style loss are defined between the feature map of generator and semantic label map of discriminator, respectively. KCGAN can preserve the fine-grained anatomical structure of the source domain and transfer the style of the target domain, simultaneously. In addition, this method can be easily applied with only changing the discriminator architecture and without utilizing any additional learnable or pre-trained networks. Experimental results showed that this method outperformed existing GAN-based methods in most direction of kernel conversion among three kernels.