Globally optimal segmentation of cell nuclei in fluorescence microscopy images using shape and intensity information

L Kostrykin,C Schnörr,K Rohr

DOI: https://doi.org/10.1016/j.media.2019.101536

Abstract:Accurate and efficient segmentation of cell nuclei in fluorescence microscopy images plays a key role in many biological studies. Besides coping with image noise and other imaging artifacts, the separation of touching and partially overlapping cell nuclei is a major challenge. To address this, we introduce a globally optimal model-based approach for cell nuclei segmentation which jointly exploits shape and intensity information. Our approach is based on implicitly parameterized shape models, and we propose single-object and multi-object schemes. In the single-object case, the used shape parameterization leads to convex energies which can be directly minimized without requiring approximation. The multi-object scheme is based on multiple collaborating shapes and has the advantage that prior detection of individual cell nuclei is not needed. This scheme performs joint segmentation and cluster splitting. We describe an energy minimization scheme which converges close to global optima and exploits convex optimization such that our approach does not depend on the initialization nor suffers from local energy minima. The proposed approach is robust and computationally efficient. In contrast, previous shape-based approaches for cell segmentation either are computationally expensive, not globally optimal, or do not jointly exploit shape and intensity information. We successfully applied our approach to fluorescence microscopy images of five different cell types and performed a quantitative comparison with previous methods.