Detecting and tracking leukocytes in intravital video microscopy using a Hessian-based spatiotemporal approach

• Published in: Multidimensional systems and signal processing Vol. 30; no. 2; pp. 815 - 839
• Main Authors: Gregório da Silva, Bruno C., Carvalho-Tavares, Juliana, Ferrari, Ricardo J.
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2019; Springer Nature B.V
• Subjects: Annotations; Artificial Intelligence; Centroids; Circuits and Systems; Eigenvalues; Electrical Engineering; Endothelial cells; Engineering; Hessian matrices; Leukocytes; Microscopy; Signal,Image and Speech Processing; Tracking; Two dimensional analysis; Visual observation; Detecting and tracking of leukocytes; Leukocyte recruitment; Hessian-based spatiotemporal technique; Intravital video microscopy
• ISSN: 0923-6082; 1573-0824
• DOI: 10.1007/s11045-018-0581-5

The leukocyte recruitment analysis is an important step to understand the interactions between leukocytes and endothelial cells in the microcirculation of living animals. Performed preferably by the intravital video microscopy technique, this procedure usually requires an expert visual analysis, which is prone to the inter- and intra-observer variability. Such problem claims, therefore, an automated method to detect and track these cells. To this end, we developed an approach that combines two different analyses: in the first (2D), all video frames are individually processed by using a blob-like structure detector to find the leukocyte centroids, while in the second (2D + t), a spatiotemporal image (created by stacking all video frames) is processed by a tubular-like structure detector, which is used to determine the leukocyte trajectories over time. For both analyses, the detectors are based on the relationship between Hessian matrix eigenvalues locally obtained from image sequences. Evaluation of the proposed approach was conducted by comparing our technique to the manual annotations using precision, recall and F 1 -score measures in two video sequences. The average results for these measures were, respectively, 0.84, 0.64, and 0.72 for the first video, and 0.84, 0.87, and 0.86 for the second. These results suggested that our proposed approach is comparable with manual annotations performed by the experts and has an excellent potential for use in real circumstances. Moreover, it can reduce the observer variabilities and the burden for visual analysis.