Touching adipocyte cells decomposition using combinatorial optimization

Cell detection plays a significant role in automated biomedical image analysis. However, it is challenging to achieve accurate detection due to dense crowding/touching of cells. In this paper, we propose a robust decomposition algorithm for cell detection on adipocyte images. It formulates the decomposition into a cut selection problem using the polygon triangulation approximation and a specific-defined concavity measurement, and utilizes combinatorial optimization to select the optimal cuts. The proposed algorithm can effectively handle contour noises, large shape variance and size difference. In addition, the decomposition seeds preserve original cell shapes to facilitate the subsequent segmentation. The algorithm is well tested with 231 adipocyte microscopic image patches, which contain 2-17 touching cells with a variety of morphological complexity. The comparative experiments with the recent state of the arts demonstrate the superior performance of the proposed method.