Pancreatic Neoplasm Image Translation Based on Feature Correlation Analysis of Cross-Phase Image

CT arterial phase images can provide a powerful auxiliary to formulate pancreatic neoplasm diagnosis and treatment plans. In the absence of such images, we can use the image translation model convert CT images of other phases into CT arterial phase synthetic images. Under the supervision of manual labeling by experts or pixel-level labeling, the model can achieve better performance. However, for pancreatic neoplasm image translation, such labels are usually scarce. In this regard, we use the easily obtained paired but unaligned cross-phase real pancreatic neoplasm images as labels and constructs a cross-phase image feature correlation analysis-based image translation method (CFCA-IT). This method analyzes the image feature correlation between the synthetic images and real images and takes it as the training constraint of the translation model. Simulation experiments show that CFCA-IT can further improve the translation performance of the five state-of-the-art translation models.