Contrast CT image generation model using CT image of PET/CT

Abstract

Contrast CT imaging is a useful method to highlight structures such as blood vessels and tissue. The aim of this study was to generate contrast enhanced CT image for defining region of blood vessels and tumor from PET/CT image using deep learning approach. The deep learning approach in this study was based on conditional generative adversarial network (cGAN). The structure of cGAN composed of two convolutional neural network (CNN), generator and discriminator. Generator was trained to generate contrast CT (G-CCT) image from CT image which cannot be discriminated from the original contrast CT (CCT) image. Discriminator was trained to discriminate CCT image from G-CCT image by generator. To improve the generation performance of the model, we applied adaptive histogram equalization as image preprocessing. We compare the model trained with CT and the model with histogram equalized CT. Structural similarity index measurement (SSIM) was calculated to measure the similarity between G-CCT and CCT image. We defined and compared the region of lymphoma in left upper cervical lymph node level 2 using CT, CCT, and G-CCT. Detect the lymph node cancer was calculated by short to long axis ratio (S/L) method. In the case of the model trained with CT, SSIM was 0.9493±0.0183 in training phase and 0.9055±0.0484 in testing phase. In the case of the model trained with histogram equalized CT, SSIM was 0.954±0.0149 in training phase and 0.9081±0.047 in testing phase. G-CCT images showed more enhanced contrast in blood vessels and tumor region than CT image. Lymph node cancer S/L of CT, CCT, and G-CCT were 0.412, 0.395, and 0.397, respectively. The tumor S/L of generated contrast CT image was evaluated similar to these of real contrast CT image. This contrast CT image generation based on deep learning is helped for a cost-effective and less-hazardous process of acquiring contrast CT image to patients as well as more anatomical information with only CT scan.