Segmentation of cell nuclei. The number and morphology of cell nuclei reveals essential information for diagnosing breast cancer and for selecting the most effective therapy. The size and shape of tumor nuclei are considered for determining the tumor grade in Hematoxylin-and-Eosin-stained (H&E) images. The percentage of positively-stained cells is quantified in immunohistochemically-stained (IHC) images in order to determine the ER or PR receptor status. The manual assessment of cell nuclei is time consuming and often subject to inter- and intra- observer variabilty. In task 5.3, we have therefore developed a novel image analysis method for the automatic detection of cell nuclei in histological images. The extracted information provides the basis for the computation of derived quantitative features in task 5.4. A paper on the developed analysis method was submitted to the SPIE Medical Imaging 2016 conference. A major challenge in the automatic detection of nuclei in histological images are clusters of aggregated nuclei. Many existing software solutions struggle to separate the individual nuclei in such clusters and can, thereby, produce severe detection errors. Our methods tackles this problem through a multi-step detection algorithm. Our method starts with the computation of a nucleus probability map. The pixels in this map represent the likelihood for a nucleus being present at that particular position. The probability map is computed by a machine-learning algorithm that takes statistics on the local stain intensities into account in order to determine the corresponding probability value. The stain intensities are considered at different scales. In this manner, the algorithm integrates information from the local surrounding of every pixel and enables the distinction between nuclei centers and periphery. After the computation of the nucleus probability map, an optimized region extraction and merging algorithm is applied in order to segment the individual nuclei. This algorithm first identifies connected regions of the highest probability values. Afterwards, the algorithm iteratively merges neighboring regions of lower probability values as long as a specific merging condition is met. The merging condition is set with respect to the minimum size of the nuclei to detect. Finally, as shown in Fig. 3, the center points of the merged regions are identified as the nuclei centers.
