A novel concavity based method for automatic segmentation of touching cells in microfluidic chips. (15th September 2022)

• Record Type: Journal Article
• Authors: Zhang, Zhonghao; Li, Qiqiang; Song, Wen; Wei, Pengfei; Guo, Jing
• Is Part Of: Expert systems with applications. Volume 202(2022)
• Publication Date: 2022-09-15
• Subjects: Deep learning -- Image processing -- Touching cells segmentation -- Microfluidic chips; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.117432
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Microfluidic systems have important application value in biology and medicine research. However, a major challenge in Microfluidic based cell analysis is to automatically segment touching cells in the microscopic images. In this paper, we propose a novel automatic cell segment method based on concave point detection and matching. On the basis of high-quality image deblurring, we adopt a UNet++ based deep neural network to accurately extract cell contours by taking as input both bright and dark-field images, and trained with a Dice loss based objective function. Then, we propose a method that extracts concave points from the contours by detecting the convex hull defects, and the issue of missing concave points is addressed by considering the distance between the starting and the ending point of defect. Finally, to overcome the limitation of existing methods that the accuracy of segmentation is highly dependent on the accuracy of concave point detection, we propose a concave points matching condition based on compactness to obtain the concave point pairs for segmentation. Experimental results show that our method can effectively segment touching cells with high accuracy, and is robust to different cell concentration levels. Highlights: Propose a novel deep learning based cell segmentation method for microfluidic chips. Both the bright and dark field information is used in contour extraction. Novel concave point detection and matching algorithms are proposed. Achieve over 90% accuracy on most experiments of different concentration levels.