Wenqing Sun1, Tzu-Liang Bill Tseng2, Jianying Zhang3, Wei Qian4. 1. Department of Electrical and Computer Engineering, University of Texas at El Paso, 500 West University Avenue, El Paso, TX 79968, USA. 2. Department of Industrial, Manufacturing and Systems Engineering, University of Texas at El Paso, 500 West University Avenue, El Paso, TX 79968, USA. 3. Department of Biological Sciences, University of Texas at El Paso, 500 West University Avenue, El Paso, TX 79968, USA; Sino-Dutch Biomedical and Information Engineering School, Northeastern University, No.11, Lane 3, Wenhua Road, Heping District, Shenyang, Liaoning 110819, China. 4. Department of Electrical and Computer Engineering, University of Texas at El Paso, 500 West University Avenue, El Paso, TX 79968, USA; Sino-Dutch Biomedical and Information Engineering School, Northeastern University, No.11, Lane 3, Wenhua Road, Heping District, Shenyang, Liaoning 110819, China. Electronic address: wqian@utep.edu.
Abstract
BACKGROUND AND OBJECTIVE: A large number of labeled medical image data is usually a requirement to train a well-performed computer-aided detection (CAD) system. But the process of data labeling is time consuming, and potential ethical and logistical problems may also present complications. As a result, incorporating unlabeled data into CAD system can be a feasible way to combat these obstacles. METHODS: In this study we developed a three stage semi-supervised learning (SSL) scheme that combines a small amount of labeled data and larger amount of unlabeled data. The scheme was modified on our existing CAD system using the following three stages: data weighing, feature selection, and newly proposed dividing co-training data labeling algorithm. Global density asymmetry features were incorporated to the feature pool to reduce the false positive rate. Area under the curve (AUC) and accuracy were computed using 10 fold cross validation method to evaluate the performance of our CAD system. The image dataset includes mammograms from 400 women who underwent routine screening examinations, and each pair contains either two cranio-caudal (CC) or two mediolateral-oblique (MLO) view mammograms from the right and the left breasts. From these mammograms 512 regions were extracted and used in this study, and among them 90 regions were treated as labeled while the rest were treated as unlabeled. RESULTS: Using our proposed scheme, the highest AUC observed in our research was 0.841, which included the 90 labeled data and all the unlabeled data. It was 7.4% higher than using labeled data only. With the increasing amount of labeled data, AUC difference between using mixed data and using labeled data only reached its peak when the amount of labeled data was around 60. CONCLUSIONS: This study demonstrated that our proposed three stage semi-supervised learning can improve the CAD performance by incorporating unlabeled data. Using unlabeled data is promising in computerized cancer research and may have a significant impact for future CAD system applications.
BACKGROUND AND OBJECTIVE: A large number of labeled medical image data is usually a requirement to train a well-performed computer-aided detection (CAD) system. But the process of data labeling is time consuming, and potential ethical and logistical problems may also present complications. As a result, incorporating unlabeled data into CAD system can be a feasible way to combat these obstacles. METHODS: In this study we developed a three stage semi-supervised learning (SSL) scheme that combines a small amount of labeled data and larger amount of unlabeled data. The scheme was modified on our existing CAD system using the following three stages: data weighing, feature selection, and newly proposed dividing co-training data labeling algorithm. Global density asymmetry features were incorporated to the feature pool to reduce the false positive rate. Area under the curve (AUC) and accuracy were computed using 10 fold cross validation method to evaluate the performance of our CAD system. The image dataset includes mammograms from 400 women who underwent routine screening examinations, and each pair contains either two cranio-caudal (CC) or two mediolateral-oblique (MLO) view mammograms from the right and the left breasts. From these mammograms 512 regions were extracted and used in this study, and among them 90 regions were treated as labeled while the rest were treated as unlabeled. RESULTS: Using our proposed scheme, the highest AUC observed in our research was 0.841, which included the 90 labeled data and all the unlabeled data. It was 7.4% higher than using labeled data only. With the increasing amount of labeled data, AUC difference between using mixed data and using labeled data only reached its peak when the amount of labeled data was around 60. CONCLUSIONS: This study demonstrated that our proposed three stage semi-supervised learning can improve the CAD performance by incorporating unlabeled data. Using unlabeled data is promising in computerized cancer research and may have a significant impact for future CAD system applications.