An Efficient Multi-center Abdominal Organ Segmentation Network

Wang Chong

Keywords: Semi-supervised · Multi-center · Abdominal Organ.

Abstract: Efficient abdominal organ segmentation has important clinical significance,but the challenges in label complexity and label professional is extremely limiting this task. Semi-supervised learning (SSL) has impressively improved the label efficiency on medical image segmentation tasks with numerous unlabeled images, but in multi-center situation, its extremely few labeled data enlarges the instability of these methods. In this paper, we designed a new network structure to address multi-organ segmentation from abdominal multi-center and multi-disease CT examinations. To deal with clinical requirements and obtain a low GPU memory but still efficient deep learning model, we use UNeXt which is a Convolutional multilayer perceptron (MLP) based network for image segmentation to achieve abdominal multi-organ segmentation. The network uses tokenized MLPs in latent space reduces the number of parameters and computational complexity while being able to result in a better representation to help segmentation. To ensure the stability of semi-supervised learning in multi-center data, we use an advanced self-training framework(namely ST++). The framework selective retraining via prioritizes reliable unlabeled images based on holistic prediction-level stability. The network structure we use combined with advanced self-training strategy can solve the problem of multi-center data instability in semi-supervised learning and achieve good segmentation effect at the same time. Our experiments show that our method gives strong results on the Dice similarity coefficient, especially for liver and kidney segmentation, and does not require an excessively long inference time.