A Semi-Supervised Multi-Organ Segmentation Method via Cross Teaching between CNN and Transformer

Ru Wang

Abstract: The automatic segmentation of abdominal CT multi -organs can improve the efficiency of clinical work processes such as disease diagnosis, prognosis analysis, and treatment plan. However, for medical images, the acquisition of data is usually expensive, because it requires professional knowledge and time to generate accurate annotations. We proposed a cross teaching semi-supervised medical image segmentation model based on CNN and Transformer. At the same time, two deep neural networks were trained, and their mutual teaching combined their respective learning paradigms to improve model performance. The segmentation of the experiment on the data shows that our model is effective. Our experiment show that the the average running time spend on one data is 719.9262(s), maximum GPU memory required in our experiment is 269, average area under GPU memory time curve is 191794.82, and average area under GPU memory time curve is 36302.29.