Pancreatic Image Augmentation Based on Local Region Texture Synthesis for Tumor Segmentation

High-accuracy segmentation of lesions in pancreatic images is essential for computer-aided precision diagnosis and treatment. The segmentation accuracy of deep learning-based segmentation models depends on the number of annotated pancreatic tumor images. Due to the high cost of labeling, the size of the training set for segmentation models is usually small. This paper proposes an image augmentation model based on local region texture generation. For pancreas images (background) and tumor images (foreground) with ablated regions, the model can generate image textures for the remaining blank areas after combining the two images to obtain new samples. To improve the texture continuity between the tumor region and surrounding tissues in the generated image, this paper constructs a three-level loss function to constrain the training of the augmented model. Simulation experiments on the pancreatic tumor image set provided by the partner hospital show that the Dice coefficient of the segmentation model trained on the dataset augmented by the proposed model improves by 2.4% compared with the current optimal method when the number of real images is sparse, which proves its effectiveness and feasibility.