A boundary evidence controlled level set inference method for nuclei instance segmentation in histopathology images

Periodic B-spline (PBS) has been a successfully used technique in histopathology image segmentation. However, it is limited when dealing with multi-instance objects in real-world biomedical applications. Moreover, its segmentation results are quite sensitive to the initial settings of the model and highly depend on the selection of control points. To address these issues and boost the canonical PBS to a new level of adaptive fitting, we present Level-Set B-Spline (LSBS), a novel definition of nuclear contour inference that employs B-splines under the energy minimization of a variational level set functional, which can promote each other. To this end, we adopt a two-stage approach: The first stage solves the marker detection and image segmentation problem and the second stage solves the joint modeling of the missing contour and the latent shape. To do so, we propose the use of modified concavity measurement and optimal boundary-to-marker association as an efficient basis for computing a periodic set of knots to drive the LSBS. For any probe image, our approach enables to analysis of free-lying and overlapping cell nuclei with weak boundaries. We demonstrated the superiority of the proposed LSBS by evaluating it on various datasets, including the challenging TCGA KIRC and Kumar nuclei, where the consistently state-of-the-art performances were achieved for comparison against state-of-the-art image analysis-based methods.