Computerized detection of leukocytes in microscopic leukorrhea images.
Jing Zhang,Ya Zhong,Xiangzhou Wang,Guangming Ni,Xiaohui Du,Juanxiu Liu,Lin Liu,Yong Liu
DOI: https://doi.org/10.1002/mp.12381
IF: 4.506
2017-01-01
Medical Physics
Abstract:Purpose: Detection of leukocytes is critical for the routine leukorrhea exam, which is widely used in gynecological examinations. An elevated vaginal leukocyte count in women with bacterial vaginosis is a strong predictor of vaginal or cervical infections. In the routine leukorrhea exam, the counting of leukocytes is primarily performed by manual techniques. However, the viewing and counting of leukocytes from multiple high-power viewing fields on a glass slide under a microscope leads to subjectivity, low efficiency, and low accuracy. To date, many biological cells in stool, blood, and breast cancer have been studied to realize computerized detection; however, the detection of leukocytes in microscopic leukorrhea images has not been studied. Thus, there is an increasing need for computerized detection of leukocytes. Methods: There are two key processes in the computerized detection of leukocytes in digital image processing. One is segmentation; the other is intelligent classification. In this paper, we propose a combined ensemble to detect leukocytes in the microscopic leukorrhea image. After image segmentation and selecting likely leukocyte subimages, we obtain the leukocyte candidates. Then, for intelligent classification, we adopt two methods: feature extraction and classification by a support vector machine (SVM); applying a modified convolutional neural network (CNN) to the larger subimages. If different methods classify a candidate in the same category, the process is finished. If not, the outputs of the methods are provided to a classifier to further classify the candidate. Results: After acquiring leukocyte candidates, we attempted three methods to perform classification. The first approach using features and SVM achieved 88% sensitivity, 97% specificity, and 92.5% accuracy. The second method using CNN achieved 95% sensitivity, 84% specificity, and 89.5% accuracy. Then, in the combination approach, we achieved 92% sensitivity, 95% specificity, and 93.5% accuracy. Finally, the images with marked and counted leukocytes were obtained. Conclusion: A novel computerized detection system was developed for automated detection of leukocytes in microscopic images. Different methods resulted in comparable overall qualities by enabling computerized detection of leukocytes. The proposed approach further improved the performance. This preliminary study proves the feasibility of computerized detection of leukocytes in clinical use. (C) 2017 American Association of Physicists in Medicine