Simultaneous Peritoneal Tumor Detection Algorithm based on Domain Adaptation

Abstract

At present, most of the work is based on deep neural network to construct simultaneous peritoneal tumor detection algorithm. The prerequisite for the successful application of these algorithms is that the training set and the test set are independent and identically distributed, that is, the algorithm needs a large number of training samples with the same distribution as the target application. In order to effectively use the public data set with sufficient data to assist the training, and to get the model with superior performance index even when the data amount is limited, we propose a simultaneous peritoneal tumor detection algorithm based on domain adaptation. Specifically, we realize edge distribution alignment based on covariance matrix, and propose two constraints based on feature space optimization and conditional distribution alignment, so that the algorithm can effectively transfer knowledge by using data sets with the same tasks but different distributions. The model can learn the interface fitting to the specific data set even if there is only a small amount of labeled data. Extensive experiments show that the proposed algorithm based on domain adaptation can significantly improve the recognition performance of the model.