One-to-multi-site Unsupervised Domain Expansion For Kidney Segmentation In Abdominal CTA Images

As one of the cancers with the most incidence,the first step in performing nephron sparing surgery(NSS)is to scan the patient’s abdominal CT scan using computed tomographyenhanced radiography to segment the kidney and tumor tissue from the CTA image.The 3D model is used to determine the location,size,and morphological characteristics of the lesion to provide doctors with visual guidance,which is particularly important for the formulation of NSS surgical plans.With the development of a number of international fully-supervised kidney semantic segmentation competitions,the relevant indicators of the deep learning automated segmentation results can reach above 90% of the manual annotation.However,it is difficult to directly apply the model trained on labeled data from one hospital to the multi-site target domain data from other hospitals,and labeling these data will consume a lot of manpower and time.In response to this problem,the Unsupervised Domain Adaptation(UDA)method was proposed.The UDA method aims to transfer the application domain of the model from the source domain to the unlabeled target domain using the existing labeled data so as to achieve a model effect close to the fully-supervised training method,which greatly saves the cost of labeling target domain data.However,medical images are usually collected from multiple medical institutions and scanning sites.The traditional UDA methods only consider the target domain constituted by a single data center,and furthermore,do not pay attention to the performance of the model in the source domain,which limits the application scenarios and performance of the model.Therefore,there is an urgent need for a one-to-multi-site Unsupervised Domain Expansion method that uses unlabeled data from multiple sites to enhance the generalization ability of the model on target domain data and observe no performance degradation on original source domain data.The main research contents of this paper are as follows:(1)Starting from the convolutional neural network,this paper proposes an entropy-based Hybrid Uncertainty Learning(HUL)method to make the model adaptively generalize to the target domain data without additional annotation.HUL uses 3D U-Net as the main architecture of the segmentation network to reduce the uncertainty of the model to the target domain image from two complementary levels to reduce the inter-domain gap.First of all,this paper proposes a 3D entropy minimization loss function,based on the model’s false segmentation results in the target domain image accompanied by the observation of high entropy values,and performs pixel-wise uncertainty entropy minimization training.Secondly,in order to learn the morphological characteristics of the lesion tissue,we built a 3D discriminant network,and the segmentation network used as generator for adversarial learning.The discriminant network takes the entropy map of the output result from the segmentation network as input which comes from the source domain or the target domain.To distinguish,force the segmentation network from a global level to produce a segmentation result close to the source domain image.(2)However,for unlabeled target domain data from multiple sites,heterogeneity appear between different sites,making the adaptation difficulty between centers vary from each other,and HUL is not capable of paying attention to the difference between sites(Inter-site gap)and cannot maintain the performance of the model on the source domain.In response to these two problems,this paper proposes a Dynamic Credible Sample Strategy(DCSS),which further enhances the generalization performance of the model on multi-site data by reducing the intersite gap.Combining the entropy map generated by the model in the target domain image and the post-processing prediction result,the weighted lesion mean entropy represents the confidence of the model for the image.DCSS adds the pseudo-labels of the cases with the highest confidence to the fully supervised source domain data pool.The remaining untrusted cases in the target domain are used as the new target domain,and multiple rounds of adaptive learning using HUL are performed.In the experiment of target domain,the final average dice coefficient of kidney and tumor is 83.8%,which reaches more than 90% of the state-of-art fullysupervised performance in Ki TS19 segmentation challenge.In the source domain,dice increased by 0.6% on average after UDE,maintaining or even improving the performance on the source domain.The proposed HUL and DCSS have been verified in both the source domain and the target domain with decent performance in this one-to-multi-site unsupervised domain expansion task of kidney CT which proves that our method could play a visual guiding role in practice while saving labels.