In this paper, we propose a novel method, an adaptive localizing region-based level set using convolutional neural network, for improving performance of maxillary sinus segmentation. The healthy sinus without lesion inside is easy for conventional algorithms. However, in practice, most of the cases are filled with lesions of great heterogeneity which lead to lower accuracy. Therefore, we provide a strategy to avoid active contour from being trapped into a nontarget area. First, features of lesion and maxillary sinus are studied using a convolutional neural network (CNN) with two convolutional and three fully connected layers in architecture. In addition, outputs of CNN are devised to evaluate possibilities of zero level set location close to lesion or not. Finally, the method estimates stable points on the contour by an interactive process. If it locates in the lesion, the point needs to be paid a certain speed compensation based on the value of possibility via CNN, assisting itself to escape from the local minima. If not, the point preserves current status till convergence. Capabilities of our method have been demonstrated on a dataset of 200 CT images with possible lesions. To illustrate the strength of our method, we evaluated it against state-of-the-art methods, FLS and CRF-FCN. For all cases, our method, as assessed by Dice similarity coefficients, performed significantly better compared with currently available methods and obtained a significant Dice improvement, 0.25 than FLS and 0.12 than CRF-FCN, respectively, on an average.
Representation of language is the rst and critical task for Natural Language Understanding (NLU) in a dialogue system. Pretraining, embedding model, and ne-tuning for intent classi cation and slot-lling are popular and well-performing approaches but are time consuming and ine cient for low-resource languages. Concretely, the out-of-vocabulary and transferring to di erent languages are two tough challenges for multilingual pretrained and cross-lingual transferring models. Furthermore, quality-proved parallel data are necessary for the current frameworks. Stepping over these challenges, di erent from the existing solutions, we propose a novel approach, the Hypergraph Transfer Encoding Network "HGTransEnNet. e proposed model leverages o -the-shelf high-quality pretrained word embedding models of resource-rich languages to learn the high-order semantic representation of low-resource languages in a transductive clustering manner of hypergraph modeling, which does not need parallel data. e experiments show that the representations learned by "HGTransEnNet" for low-resource language are more e ective than the state-of-the-art language models, which are pretrained on a large-scale multilingual or monolingual corpus, in intent classi cation and slot-lling tasks on Indonesian and English datasets.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.