Line Junction Detection Without Prior-Delineation of Curvilinear Structure in Biomedical Images

The line junction detection is a fundamental step in many computer vision applications, especially in biomedical image analysis. Most of the existing studies determine the junction position after delineating curvilinear structure, thus the detection accuracy relies heavily on the previous steps for curvilinear extraction, such as image segmentation and skeletonization. In this paper, we treat the detection of line junctions as an independent task without prior knowledge of curvilinear structures. We present the mathematical definition and properties of line junctions, and propose a new method called Junction Recognition (JUNR). It first maps the raw images into score matrices (or called score images) by the measurements based on line junction properties, then detects and screens blobs from the score images for identifying the regions covering junction points. Finally, it refines the locations of line junctions as well as their branch properties. A distinct advantage of JUNR is that it can be directly applied to raw images without knowing curvilinear structure beforehand. Besides, since JUNR is a rule-based method, it requires no training data and avoids the labor-intensive labeling work. We conducted experiments on two typical kinds of biomedical images, including both simulated and real images with curvilinear structures. Both qualitative and quantitative results demonstrate its good performance for junction detection and characterization.