AVP - Research and Impact

ADDRESS
WEBSITE
1. Chromosome Classification with Convolutional Neural Network based Deep Learning

Author:Zhang, WB;Song, SF;Bai, TM;Zhao, YX;Ma, F;Su, JL;Yu, LM

Source:2018 11TH INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING, BIOMEDICAL ENGINEERING AND INFORMATICS (CISP-BMEI 2018),2018,Vol.

Abstract:Karyotyping plays a crucial role in genetic disorder diagnosis. Currently Karyotyping requires considerable manual efforts, domain expertise and experience, and is very time consuming. Automating the karyotyping process has been an important and popular task. This study focuses on classification of chromosomes into 23 types, a step towards fully automatic karyotyping. This study proposes a convolutional neural network (CNN) based deep learning network to automatically classify chromosomes. The proposed method was trained and tested on a dataset containing 10304 chromosome images, and was further tested on a dataset containing 4830 chromosomes. The proposed method achieved an accuracy of 92.5%%, outperforming three other methods appeared in the literature. To investigate how applicable the proposed method is to the doctors, a metric named proportion of well classified karyotype was also designed. An result of 91.3%% was achieved on this metric, indicating that the proposed classification method could be used to aid doctors in genetic disorder diagnosis.
2. Length-of-Stay Prediction for Pediatric Patients With Respiratory Diseases Using Decision Tree Methods

Author:Ma, F;Yu, LM;Ye, LS;Yao, DD;Zhuang, WF

Source:IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS,2020,Vol.24

Abstract:Accurate prediction of a patient's length-of-stay (LOS) in the hospital enables an efficient and effective management of hospital beds. This paper studies LOS prediction for pediatric patients with respiratory diseases using three decision tree methods: Bagging, Adaboost, and Random forest. A data set of 11,206 records retrieved from the hospital information system is used for analysis after preprocessing and transformation through a computation and an expansion method. Two tests, namely bisection test and periodic test, are designed to assess the performance of the prediction methods. Bagging shows the best result on the bisection test (0.296 RMSE, 0.831 R-2, and 0.723 Acc +/- 1) for the testing set of the whole data test. The performances of the three methods are similar on the periodic test, whereas Adaboost performs slightly better than the other two methods. Results indicate that the three methods are all effective for the LOS prediction. This study also investigates the importance of different data fields to the LOS prediction, and finds that hospital treatment-related data fields contribute more to the LOS prediction than other categories of fields.
3. Prediction of Days in Hospital for Children Using Random Forest

Author:Wang, CG;Dong, XL;Yu, LM;Ye, LS;Zhuang, WF;Ma, F

Source:2017 10TH INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING, BIOMEDICAL ENGINEERING AND INFORMATICS (CISP-BMEI),2017,Vol.2018-January

Abstract:In this study, a method was developed to predict the number of hospitalization days of infant patients. The random forest algorithm, along with a data set consisted by records extracted from a hospital information system, was utilized to develop a model to predict the days in hospital. When half of randomly selected records was used as training set to train the random forest algorithm and the other half was used as testing set to test the trained model, the random forest method achieved good predictive accuracy with RMSE being 0.314, R-2 being 0.706, IR1 being 0.545, and Acc 1 being 71%%, which is better than the results obtained by Adaboost method and Bagging method. Experiment on three subgroups of records: a group with all data, a group with records having less than or equal to 14 days in hospital, and a group with records having greater than 14 days in hospital, shows that the prediction of the developed method on the group having more than 14 days in hospital was better than predictions on other groups. Analysis to the importance of three different types of feature sets to the accuracy of prediction reveals that the feature set relating to personal information contribute more to the prediction than other types of features.
4. Temporal change analysis for computer aided mass detection in mammography

Author:Ma, F;Yu, LM;Liu, G;Niu, Q

Source:2014 7TH INTERNATIONAL CONFERENCE ON BIOMEDICAL ENGINEERING AND INFORMATICS (BMEI 2014),2014,Vol.

Abstract:This paper presents a method to extract change information from temporal mammogram pairs and to incorporate the temporal change information in the malignant mass classification. In this method, a temporal mammogram registration framework which is based on spatial relations between regions of interest and graph matching was used to create correspondences between regions of current mammogram and regions of previous mammogram. 18 image features were then used to capture the differences (temporal changes) between the matched regions. To assess the contribution of temporl change information to the mass detection, 4 methods were designed to combine mass classification on image features measured on single regions and mass classification on temporal features to improve overall mass classification. The method was tested on 95 pairs of temporal mammograms using k-fold cross validation procedure. The experimental results showed that, when combining two classification results using linear combination or by taking minimum value, the A(z) score of overall classification performance increased from 0.8843 to 0.8958 and 0.8962 respectively. The results demonstrated that registering temporal mammograms, measuring temporal changes from matched regions and incorporating the change information in the mass classification improves the overall mass detection.
5. Rational-Orthogonal-Wavelet-Based Active Sonar Pulse and Detector Design

Author:Yu, LM;Ma, F;Lim, E;Cheng, E;White, LB

Source:IEEE JOURNAL OF OCEANIC ENGINEERING,2019,Vol.44

Abstract:A family of rational orthogonal wavelets is explored in active sonar detection to achieve Doppler robustness and noise mitigation. The broadband active sonar system is classified as wavelet-based noncoherent subband energy detection. It features a flexible broadband sonar pulse design with low peak-to-average power ratio (PAPR), and a detector of low computation complexity via fast rational wavelet filter bank. The designed family of broadband pulses is based on a combination of multiple basis functions of the complex rational orthogonal wavelet (CROW). Differing from the conventional dyadic wavelet (with an integer scale factor of 2), the CROW has a rational scale factor and much higher resolution in the partition of the time-scale plane. This family of wavelets presents great flexibility in the system design. Design example is presented with minimized PAPR to serve for applications demanding the stealth. The sonar detector is based on the CROW analysis filter bank. Detection performance is analyzed regarding its resistance to Doppler noise and multipath by simulations using a geometric underwater acoustic channel model. The performance of the wavelet-based active sonar system is compared with a linear frequency modulated pulse-based system. The receiver operating characteristic curves demonstrate that the designed system is able to cope with severe multipath, high ambient noise, and severe Doppler dispersion. The framework provides possible solutions for underwater applications in low signal-to-noise ratio and varying mobility conditions.
6. Inventory sharing strategy and optimization for reusable transport items

Author:Liu, GQ;Li, L;Chen, JH;Ma, F

Source:INTERNATIONAL JOURNAL OF PRODUCTION ECONOMICS,2020,Vol.228

Abstract:Reusable transport items (RTI), as a sustainable solution to the increasing packaging wastes generated along with the ever-growing globalized supply chains, create both challenges and opportunities to organizations for their management. A trend of organizations outsourcing their RTI activities explains the rapid emerging of a third-party RTI pooler. Yet few research studies have been conducted from the perspective of the RTI pooler for optimizing their customer allocation, and few on the decision support model for a practical case in the developing economies. This research intends to fill in this gap. Motivated by a case study of a leading RTI pooling company in China, this research proposes to implement a sharing strategy into the daily planning operation of distribution and routing. A decision support framework is developed to optimize the distribution flows and dispatching vehicle routes by the use of a two-stage solution process. Empirical results demonstrate an economic savings of 28.1%% in the transportation costs along with environmental and social advantages implicated by the shortened travel distance of vehicles.
7. Personalized Recommender Systems with Multi-source Data

Author:Wang, Yili ; Wu, Tong ; Ma, Fei ; Zhu, Shengxin

Source:Advances in Intelligent Systems and Computing,2020,Vol.1228 AISC

Abstract:Pervasive applications of personalized recommendation models aim to seek a targeted advertising strategy for business development and to provide customers with personalized suggestions for products or services based on their personal experience. Conventional approaches to recommender systems, such as Collaborative Filtering (CF), use direct user ratings without considering latent features. To overcome such a limitation, we develop a recommendation strategy based on the so-called heterogeneous information networks. This method can combine two or multiple sources datasets and thus can reveal more latent associations/features between items. Compared with the well-known ‘k Nearest Neighborhood’ model and ‘Singular Value Decomposition’ approach, the new method produces a substantial higher accuracy under the commonly used measurement which is mean absolute deviation. © 2020, Springer Nature Switzerland AG.
8. Location-aware convolutional neural networks based breast tumor detection

Author:Hu,Huafeng;Coenen,Frans;Ma,Fei;Thiyagalingam,Jeyarajan;Su,Jionglong

Source:IET Conference Publications,2018,Vol.2018

Abstract:Breast cancer is one of the most common types of cancer affecting the lives of millions. Early detection and localization of the breast cancer tissues are vital for prevention and cure. Recently, there have been a number of developments on this front, particularly in the direction of automated image analysis. Although they are instrumental in expediting the process, such approaches lack the localization information and hence still demand substantial involvement of clinicians to deliver conclusive results. In this paper, we propose a novel approach for detecting and localizing cancer tissues from mammograms. In particular, we rely on Convolutional Neural Networks for exploiting the spatial relationship of the cancer tissues for detection and localization. Our evaluations on real datasets show that the proposed method is able to classify normal and tumor tissues with the classification accuracy of 90.8%%. Furthermore, our approach achieves the sensitivity of 86.1%% in detection with 1.4 false positives per image on the localization. In comparison to the state-of-the-art approaches, our method offers an additional 1.1%% sensitivity improvement, along with reduced two false positives per image.
9. Particle Filter Based Time Series Prediction of Daily Sales of an Online Retailer

Author:Ping, XY;Chen, QY;Liu, GQ;Su, JL;Ma, F

Source:2018 11TH INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING, BIOMEDICAL ENGINEERING AND INFORMATICS (CISP-BMEI 2018),2018,Vol.

Abstract:Accurate prediction of sales is instrumental to successful management in the industries. It is crucial in formulating business strategies under uncertainties. In this paper, we consider time series in which observations are arriving sequentially. An online time series model integrating with particle filter is used for predicting sales of 80 products in a local online retailer over 400 days. We embed an Autoregressive model into a state space model and carry out time series prediction for all 80 products using a particular Particle Filter called the Sampling Importance Resampling Filter. Our experiment shows that the proposed model successfully predicts 27.5%% of sales fluctuating within 10%% of the true values. Furthermore, it outperforms the traditional Autoregressive Integrated Moving Average model by 5%% for the same metric used.
10. Re-Identification Based Automatic Matching and Annotation of Chromosome

Author:Wang, Chengyu ; Huang, Daiyun ; Guo, Jingwei ; Su, Jionglong ; Ma, Fei ; Yu, Limin

Source:Proceedings - 2019 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics, CISP-BMEI 2019,2019,Vol.

Abstract:Karyotyping of human chromosomes generally consists of three steps pre-processing, segmentation and classification. By analyzing the number and structure of chromosomes, diseases such as cancers and genetic disorders can be diagnosed. Besides the traditional methods, The Convolutional Neural Network have improved the computer vision area dramatically. When it comes to chromosome karyotyping, few research methods have been proposed to solve the problem of segmentation and classification. This paper proposes an innovative automatic strategy named Chromosome-Automatic-Annotation (CAA) model, which labels the single chromosomes in microscopic images by 1) applying a joint loss consists of softmax loss and center loss to enlarge the distance of features among the 24 classes; 2) employing the similarity matrix to annotate the single chromosome images in Query Queue with the single chromosome in Gallery Queue. With a dataset of 90624 single chromosome images, after 50 epoch training, the proposed model reached an accuracy of 98.75%% for automatic annotation of the chromosome images on a test set of 644 images. © 2019 IEEE.
11. Computer Aided Mass Detection in Mammography with Temporal Change Analysis

Author:Ma, F;Yu, LM;Liu, G;Niu, Q

Source:COMPUTER SCIENCE AND INFORMATION SYSTEMS,2015,Vol.12

Abstract:This paper presents a method to extract change information from temporal mammogram pairs and to incorporate the temporal change information in the malignant mass classification. In this method, a temporal mammogram registration framework which is based on spatial relations between regions of interest and graph matching was used to create correspondences between regions of current mammogram and regions of previous mammogram. 18 image features were then used to capture the differences (temporal changes) between the matched regions. To assess the contribution of temporal change information to the mass detection, 5 methods were designed to combine mass classification on image features measured on single regions and mass classification on temporal features to improve overall mass classification. The method was tested on 95 pairs of temporal mammograms using k-fold cross validation procedure. The experimental results showed that, when combining two classification results using linear combination or by taking minimum value, the A z score of overall classification performance increased from 0.8843 to 0.8989 and 0.8863 respectively. The results demonstrated that registering temporal mammograms, measuring temporal changes from matched regions and incorporating the change information in the mass classification improves the overall mass detection.
12. Extended ResNet and Label Feature Vector Based Chromosome Classification

Author:Wang, CY;Yu, LM;Zhu, X;Su, JL;Ma, F

Source:IEEE ACCESS,2020,Vol.8

Abstract:Human chromosome classification is essential to the clinical diagnosis of cytogenetical diseases such as genetic disorders and cancer. This process, however, is time-consuming and requires specialist knowledge. Considerable efforts have been made to automat the process. Recently, methods based on Convolutional Neural Networks achieved state-of-the-art results on the chromosome classification task. Many studies used karyotype images in performance evaluation, few studies have reported the results of human chromosome classification on microscopical images. This paper proposes a novel method to classify single chromosome images into one of 24 types. In the proposed method an extended ResNet was first devised to extract features of single chromosome images. A label feature vector was then extracted for each of 24 chromosome types based on a validation dataset. Hausdorff distance between feature vector of an input image and each of 24 label feature vectors were calculated, and the label feature vector that has minimum hausdorff distance to the feature vector of the input image was selected as the potential label of the input image. To finally allocate the single chromosomes from a same microscopical image into one of 24 types, a Label Redistribution strategy was used to shrink the label space and to increase the efficiency of chromosome classification. Experiments were implemented with 90,624 single chromosome images, 644 of which were randomly picked to form a testing set in advance. The classification accuracy on microscopical images using our proposed method achieved an accuracy of 94.72%%.
13. BROADBAND SONAR WAVEFORM DESIGN WITH RATIONAL ORTHOGONAL WAVELET

Author:Yu, LM;Ma, F;Lim, E;Cheng, E;White, LB

Source:2017 14TH INTERNATIONAL COMPUTER CONFERENCE ON WAVELET ACTIVE MEDIA TECHNOLOGY AND INFORMATION PROCESSING (ICCWAMTIP),2017,Vol.2018-February

Abstract:The paper proposes a new family of broadband sonar pulses. Each pulse is formed by a combination of orthogonal wavelet basis functions as subpulses with variable scaling factor and shifts. The wavelet adopted is a complex rational orthogonal wavelet (CROW) with a fractional scale dilation factor valued between 1 and 2. The number of subpulses, scaling factor and their time shifts are designed to ensure that the sonar pulse conforms to a designated signal spectrum. The issue of peak-to-average power ratio (PAPR) of the sonar pulse is explored and design example is presented with minimized PAPR. The proposed waveform design features efficient detection algorithm based on CROW analysis filter bank (FB). The detection performance is analyzed regarding Doppler resistance, noise mitigation and multipath by simulations. Preliminary results show that the designed detection system is able to cope with severe multipath, high ambient noise and sever Doppler dispersion.
14. Forecasting Clinical Expenditure of Child Patients Using Binary and Multi-Classification Methods

Author:Wang, CG;Pan, XY;Ye, LS;Zhuang, WS;Ma, F

Source:2018 INTERNATIONAL CONFERENCE ON BIG DATA AND ARTIFICIAL INTELLIGENCE (BDAI 2018),2018,Vol.

Abstract:In this paper, random forest algorithm (RF) and error-correction output code model (ECOC) were employed to predict the clinic expenditure of child patients with data consisting of records extracted from a hospital information system. Throughout the modelling, the training set utilized 80%% of the records selected from original data set in random and the rest of data were used in the test set. The RF received better predictive accuracy than ECOC, with RMSE being 0.152, R-2 being 0.924, vertical bar R vertical bar being 0.869, and Acc +/- 1 being 82.6%%. Additionally, RF obtained good performances on different types of charges, achieving over 80%% accuracy in average. Besides, among different types of information, clinic features gave better results, with RMSE being 0.215, R-2 being 0.844, vertical bar R vertical bar being 0.709 and Acc being 60.5%%. In comparison, the random forest generally performed better than ECOC models in most fields. To summarize, the random forest could obtain best accuracy on charge7 (Treatment Fee), with accuracy of 90.5%%, and clinic features could provide models with higher accuracy among all fields of information.
15. Nipple Detection in Mammogram Using a New Convolutional Neural Network Architecture

Author:Lin, Yuyang ; Li, Muyang ; Chen, Sirui ; Yu, Limin ; Ma, Fei

Source:Proceedings - 2019 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics, CISP-BMEI 2019,2019,Vol.

Abstract:Mammogram is an X-ray image of the breast. It plays an important role in the breast cancer early diagnosis. In recent years, computer aided detection (CAD) is used for breast cancer detection. Multi-view of mammograms are needed to achieve high accuracy of automatic detection. Since nipple is the only landmark on mammogram of different views (mediolateral oblique (MLO) and craniocaudal (CC) views), nipple detection becomes the first important step of many CAD systems. Researchers have developed different models to detect nipple in recent 20 years. Grey scale, geometric feature and breast edge's gradient are used to find the nipple on the mammogram. For most methods, MLO and CC views need to be tested separately, and obvious and subtle types of nipples also need different methods to detect. In this paper, a model with deep learning is designed to locate nipples on mammogram of both MLO and CC views. Both obvious and subtle types are used for experiment. Four convolutional neural network blocks are used to attain candidate blocks. Normalization layers are added to the proposed model in order to improve the domain adaptation. Based on the intersection of candidates, the model computes the final block of nipple. In this experiment, train set and test set are randomly attained from Digital Database for Screening Mammography (DDSM). Our proposed method achieved an overall nipple detection accuracy of 98.00%%, which outperformed three comparative methods. © 2019 IEEE.
16. Mammogram mass classification with temporal features and multiple kernel learning

Author:Ma, F;Yu, LM;Bajger, M;Bottema, MJ

Source:2015 INTERNATIONAL CONFERENCE ON DIGITAL IMAGE COMPUTING: TECHNIQUES AND APPLICATIONS (DICTA),2015,Vol.

Abstract:Based on previous work on regional temporal mammogram registration, this study investigates the combination of image features measured from single regions (single features) and image features measured from the matched regions of temporal mammograms (temporal features) for the classification of malignant masses. Three SVM kernels, the multilayer perceptron kernel, the polynomial kernel, and the gaussian radial basis function kernel, and the combination of these kernels, the multiple kernel learning method, were applied to both single and temporal features for the mass classification. To combine the two types of features, 3 combination rules, Linear combination, Max and Min, were used to combine classification results obtained on single and temporal features. The results showed that combining the MKL classification results on single features, and MKL classification results on temporal features, with Min rule produces the best classification results. The experiment result indicates that incorporating the temporal change information in mammography mass classification can improve the performance detection.
17. Generalization and Visual Comprehension of CNN Models on Chromosome Images

Author:Wang, CY;Huang, DY;Su, JL;Yu, LM;Ma, F

Source:2020 4TH INTERNATIONAL CONFERENCE ON CONTROL ENGINEERING AND ARTIFICIAL INTELLIGENCE (CCEAI 2020),2020,Vol.1487

Abstract:Computer-aided image classification has achieved start-of-the-art performance since Convolutional Neural Network structures were employed. Classical neural networks such as AlexNet and VGG-Net inspired several rules of designing network models. Besides benchmark datasets such as MNIST, CIFAR and ImageNet, classification performance of medical images such as chromosome karyotyping images also improved via Convolutional Neural Network. However, there are few studies on generalization among different datasets. In this paper, we designed a neural network with nine layers, and achieved classification accuracy of 0.984, 0.816 and 0.921 on the dataset of MNIST, CIFAR and chromosome karyotype images. We also visualized the output of several layers of the model and explained that smooth output between neural network layers may induce lower accuracy on classification.
18. ECG Signal Classification with Deep Learning for Heart Disease Identification

Author:Zhang, WB;Yu, LM;Ye, LS;Zhuang, WF;Ma, F

Source:2018 INTERNATIONAL CONFERENCE ON BIG DATA AND ARTIFICIAL INTELLIGENCE (BDAI 2018),2018,Vol.

Abstract:Electrocardiogram (ECG) signal is widely used in medical diagnosis of heart diseases. Automatic extraction of relevant and reliable information from ECG signals has not been an easy task for computerized system. This study proposes to use 12-layer 1-d CNN to classify 1 lead individual heartbeat signal into five classes of heart diseases. The proposed method was tested on MIT/BIH arrhythmia database and results were measured using positive predictive value, sensitivity and F1 score. Our proposed method obtained a positive predictive value of 0.977, sensitivity of 0.976, and F1 score of 0.976. Comparing with the results obtained by other four methods on the same database, our method was found superior on all three measures.
19. Online Shop Daily Sale Prediction Using Adaptive Network-Based Fuzzy Inference System

Author:Liang, Yuanbang ; Jia, Yunyu ; Li, Jinglin ; Chen, Meiyi ; Hu, Yifan ; Shi, Yinan ; Ma, Fei

Source:Proceedings - 2019 12th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics, CISP-BMEI 2019,2019,Vol.

Abstract:Online shopping is an increasingly popular way of purchasing among customers. Demands from online shopping are usually highly dynamic. Inventory management, to many online shop owners, hence is a challenging task. Accurate forecasting of daily sales helps better replenishment and sale. This paper proposes a method for online shop daily sale forecasting. In the method, Kalman Filter was firstly applied on the historic sale data to smooth the data. Adaptive Network-based Fuzzy Inference System (ANFIS) was then built to achieve time series forecasting. Sale histories of an online shop was used to evaluate the method. Overall performance of ANFIS was evaluated with MAPE and ACC({pm}), and were found to be 1.0725 and 40%%, respectively. The results were found to be slightly better than Back Propagation Neural Network which has MAPE 4.9891 and ACC({\pm}) 39.85{\%%}. 200 products were divided into 6 groups based on corresponding MAPE and ACC({\pm}). The results showed that daily sales of 22.86%% products were predicted accurately, while that of 34.29%% products were badly predicted. The results also indicated that membership function inserted in the model causes little impact on prediction accuracy. It was tested that the average performance indicator ACC({\pm}) ({\%%}) stays stable with different membership functions. © 2019 IEEE.
20. Kalman Filter Based Time Series Prediction of Cake Factory Daily Sale

Author:Wu, JX;Fang, Q;Xu, YY;Su, JL;Ma, F

Source:2017 10TH INTERNATIONAL CONGRESS ON IMAGE AND SIGNAL PROCESSING, BIOMEDICAL ENGINEERING AND INFORMATICS (CISP-BMEI),2017,Vol.2018-January

Abstract:Accurate prediction of future daily sales is a crucial step towards optimal management of daily production of a cake factory. In this study, an interacting multiple model integrated kalman filter was used to predict the future daily sales of cake products. Two years daily sale history of 108 cake products were used to train and test the proposed method. Our experiments show that 1) running interacting multiple models of different orders in parallel is more effective than single classical interacting multiple model; 2) when only daily sale data was used, the proposed method predicted 33.54%% of sales within +/- 10%% of true sales; 3) when more variables, including festival and weekend, were combined into the prediction, 34.38%% of predicted sales were within +/- 10%% of true sales.
Total 26 results found
Copyright 2006-2020 © Xi'an Jiaotong-Liverpool University 苏ICP备07016150号-1 京公网安备 11010102002019号