基于BERT的电子病历命名实体识别

电子病历是保存、管理、传输病人医疗记录的重要资源,是医生诊治疾病的重要文本记录。通过电子病历命名实体识别(NER)技术能够高效、智能地从电子病历中抽取症状、疾病、药名等诊疗信息,有利于结构化电子病历,使之能够使用机器学习等技术进行诊疗规律挖掘。为了高效识别电子病历中的命名实体,提出一种融合对抗训练(FGM)的基于BERT与双向长短期记忆网络(BILSTM)的命名实体识别方法(BERT-BILSTM-CRF-FGM,BBCF),对2017全国知识图谱与语义计算大会(CCKS2017)提供的中文电子病历语料做修正等预处理后,采用BBCF模型识别该语料中5种实体的平均F1值为92.84%,比基于膨胀卷积网络的BERT模型(BERT-IDCNN-CRF)和基于BILSTM的条件随机场模型(BILSTM-CRF)有更高的F1值和更快的收敛速度,能够更加高效地结构化电子病历文本。

多模态医疗数据中海量小文件存储优化方法

Hadoop分布式文件系统(HDFS)通常用于大文件的存储和管理,当进行海量小文件的存储和计算时,会消耗大量的NameNode内存和访问时间,成为制约HDFS性能的一个重要因素.针对多模态医疗数据中海量小文件问题,提出一种基于双层哈希编码和HBase的海量小文件存储优化方法.在小文件合并时,使用可扩展哈希函数构建索引文件存储桶,使索引文件可以根据需要进行动态扩展,实现文件追加功能.在每个存储桶中,使用MWHC哈希函数存储每个文件索引信息在索引文件中的位置,当访问文件时,无须读取所有文件的索引信息,只需读取相应存储桶中的索引信息即可,从而能够在O(1)的时间复杂度内读取文件,提高文件查找效率.为了满足多模态医疗数据的存储需求,使用HBase存储文件索引信息,并设置标识列用于标识不同模态的医疗数据,便于对不同模态数据的存储管理,并提高文件的读取速度.为了进一步优化存储性能,建立了基于LRU的元数据预取机制,并采用LZ4压缩算法对合并文件进行压缩存储.通过对比文件存取性能、NameNode内存使用率,实验结果表明,所提出的算法与原始HDFS、HAR、MapFile、TypeStorage以及HPF小文件合并方法相比,文件读取时间更短,能够提高HDFS在处理多模态医疗数据中海量小文件时的整体性能.

基于自适应色彩聚类和上下文信息的自然场景文本检测

自然场景文本检测是图像内容分析和理解的重要前提.本文提出一种基于自适应色彩聚类和上下文信息分析的方法,用于检测自然场景图像文本.首先,将层次聚类和参数自学习策略结合,设计一种自适应色彩聚类方法,提取图像中的候选字符.该自适应色彩聚类方法能针对不同图像自动学习权重阈值,有较好的字符召回率.然后,利用文本中字符成行出现的性质,设计一种基于上下文信息的字符验证策略,既能保证较高字符召回率,也能有效移除非文本字符.最后,合并字符构建文本行,并通过后处理得到文本检测结果.在ICDAR2013公共数据集上的实验结果表明:本文分别获得74.17%的召回率,83.40%的准确率和78.52%的F得分.与其他文本检测方法相比,本文获得了较好的文本检测性能,说明本文方法的优越性.

Improved pedestrian detection with peer AdaBoost cascade

Focusing on data imbalance and intraclass variation,an improved pedestrian detection with a cascade of complex peer AdaBoost classifiers is proposed.The series of the AdaBoost classifiers are learned greedily,along with negative example mining.The complexity of classifiers in the cascade is not limited,so more negative examples are used for training.Furthermore,the cascade becomes an ensemble of strong peer classifiers,which treats intraclass variation.To locally train the AdaBoost classifiers with a high detection rate,a refining strategy is used to discard the hardest negative training examples rather than decreasing their thresholds.Using the aggregate channel feature(ACF),the method achieves miss rates of 35%and 14%on the Caltech pedestrian benchmark and Inria pedestrian dataset,respectively,which are lower than that of increasingly complex AdaBoost classifiers,i.e.,44%and 17%,respectively.Using deep features extracted by the region proposal network(RPN),the method achieves a miss rate of 10.06%on the Caltech pedestrian benchmark,which is also lower than 10.53%from the increasingly complex cascade.This study shows that the proposed method can use more negative examples to train the pedestrian detector.It outperforms the existing cascade of increasingly complex classifiers.

Chinese micro-blog sentiment classification through a novel hybrid learning model

With the rising and spreading of micro-blog, the sentiment classification of short texts has become a research hotspot. Some methods have been developed in the past decade. However, since the Chinese and English are different in language syntax, semantics and pragmatics, sentiment classification methods that are effective for English twitter may fail on Chinese micro-blog. In addition, the colloquialism and conciseness of short Chinese texts introduces additional challenges to sentiment classification. In this work, a novel hybrid learning model was proposed for sentiment classification of Chinese micro-blogs, which included two stages. In the first stage, emotional scores were calculated over the whole dataset by utilizing an improved Chinese-oriented sentiment dictionary classification method. Data with extremely high or low scores were directly labeled. In the second stage, the remaining data were labeled by using an integrated classification method based on sentiment dictionary, support vector machine(SVM) and k-nearest neighbor(KNN). An improved feature selection method was adopted to enhance the discriminative power of the selected features. The two-stage hybrid framework made the proposed method effective for sentiment classification of Chinese micro-blogs. Experiments on the COAE2014(Chinese Opinion Analysis Evaluation 2014) dataset show that the proposed method outperforms other schemes.

Detection of artificial pornographic pictures based on multiple features and tree mode

It is easy for teenagers to view pornographic pictures on social networks. Many researchers have studied the detection of real pornographic pictures, but there are few studies on those that are artificial. In this work, we studied how to detect artificial pornographic pictures, especially when they are on social networks. The whole detection process can be divided into two stages： feature selection and picture detection. In the feature selection stage, seven types of features that favour picture detection were selected. In the picture detection stage, three steps were included. 1） In order to alleviate the imbalance in the number of artificial pornographic pictures and normal ones, the training dataset of artificial pornographic pictures was expanded. Therefore, the features which were extracted from the training dataset can also be expanded too. 2） In order to reduce the time of feature extraction, a fast method which extracted features based on the proportionally scaled picture rather than the original one was proposed. 3） Three tree models were compared and a gradient boost decision tree （GBDT） was selected for the final picture detection. Three sets of experimental results show that the proposed method can achieve better recognition precision and drastically reduce the time cost of the method.

Intraretinal layer segmentation and parameter measurement in optic nerve head region through energy function of spatial-gradient continuity constraint

For the diagnosis of glaucoma,optical coherence tomography(OCT)is a noninvasive imaging technique for the assessment of retinal layers.To accurately segment intraretinal layers in an optic nerve head(ONH)region,we proposed an automatic method for the segmentation of three intraretinal layers in eye OCT scans centered on ONH.The internal limiting membrane,inner segment and outer segment,Bruch’s membrane surfaces under vascular shadows,and interaction of multiple high-reflectivity regions in the OCT image can be accurately segmented through this method.Then,we constructed a novel spatial-gradient continuity constraint,termed spatial-gradient continuity constraint,for the correction of discontinuity between adjacent image segmentation results.In our experiment,we randomly selected 20 B-scans,each annotated three retinal layers by experts.Signed distance errors of?0.80μm obtained through this method are lower than those obtained through the state-of-art method(?1.43μm).Meanwhile,the segmentation results can be used as bases for the diagnosis of glaucoma.

BGIDB:A fundus ground truth building tool with automatic DDLS classification for glaucoma research

Taking advantage of the new standard HTML5,we designed an online tool called a browser/server-based glaucoma image database builder(BGIDB)for the demarcation of the optic disk and cup’s ellipse-like boundaries.The B-spline interpolation algorithm is used,and a specially designed algorithm is proposed for classifying the disease grade according to the disc damage likelihood scale criterion,which is correlated strongly with the glaucoma process by quantity.This tool exhibits the best performance with a low overlapping error of 4.34%for the optic disk demarcation and 8.31%for the optic cup demarcation.It also has preferable time-consuming as compared to other tools and is a cross-platform system.This tool has already been utilized in building the ophthalmic image database in the cooperation of Center for Ophthalmic Imaging Research and The Second Xiangya Hospital.

Automatic segmentation of optic disc and cup for CDR calculation

Glaucoma as an irreversible blinding opioid neuropathy disease, its blindness rate is the second only after cataract in the world. The optic cup-to-disc ratio(CDR) is generally considered to be an important clinical indicator for judging the severity of glaucoma by ophthalmologists from retinal fundus image. In this letter, we propose an automatic CDR measurement method that consists of a novel optic disc localization method and a simultaneous optic disc and cup segmentation network based on the improved U shape deep convolutional neural network. Experimental results demonstrate that the proposed method can achieve superior performance when compared with other existing methods. Thus, our method can be used as a powerful tool for glaucoma-assisted diagnosis.