Comparison and Optimization Based Prior Knowledge of Models for Whole Heart Segmentation

CT angiography is a reliable and non-invasive examination. Currently, radiologists in the field of medical imaging manually segment each slice of the patient's images, which is time-consuming and labor-intensive. In this paper, we compare and optimize two artificial intelligence methods for whole-heart segmentation. On one hand, we employ the 3D U-Net, which has proven to be effective in the medical field, for training and segmentation. We optimize the data processing approach based on prior knowledge to make it more suitable for whole-heart segmentation. On the other hand, we train the UNETR, which combines the Transformer method with U-Net. We utilize data augmentation techniques such as cropping based on the positive-negative ratio to obtain improved segmentation results. Lastly, we compare the U-Net and UNETR models, and summarize and analyze their respective characteristics.